> For the Code Agent tasks among the public benchmarks above, DeepSeek-V4-Flash-0731 is evaluated with the minimal mode of DeepSeek Harness (to be released) as the agent framework
So, are they planning to announce an optimized coding agent harness as well ? DSv4 flash is a fantastic model, and my daily driver. With reasonix or pi, I can code all day long and pay a few pennies for it. No token anxiety. Whereas the same model with fireworks/openrouter, with zdr thrown in, token costs ratchet up with no explanation. Likely that the model is subsidized for gathering usage data. I am waiting for the day I can run this locally.
If you're not paying for the model through Open router, then which provider are you buying it from? I thought that if that provider was listed on open router it would be the same price as going directly to the provider? Are you buying from a provider that isn't listed on open router already?
> Are you buying from a provider that isn't listed on open router already
I am using openrouter with zdr guardrail which routes to any provider that supposedly doesnt train on user data. I also use fireworks (directly, not via openrouter) which is a provider promising zdr and has a bunch of open weights models. My issue is that these zdr providers dont transparently disclose caching/tokens etc and so they end up being far more expensive than directly using DS.
"For the Code Agent tasks among the public benchmarks above, DeepSeek-V4-Flash-0731 is evaluated with the minimal mode of DeepSeek Harness (to be released) as the agent framework, using the max reasoning effort level with temperature = 1.0, top_p = 0.95."
Somewhat relatedly, how do the economics for Huggingface work? They must be hosting petabytes of models and datasets by now. I have downloaded quite a few “just in case”, only to replace them with the later iteration months later.
Does the file hosting actually cost peanuts when you do it yourself and the cloud has shattered my understanding of what it actually costs to deliver so much data?
File hosting is pretty cheap. Egress traffic is about $90/TB in the cloud, but around $1/TB in the real world. Storage is in the realm of $5/TB/month after adding redundancy
At the scale of Huggingface, that still amounts to a lot load of money. Significantly less than if you did the same in AWS, but still a lot
That said, they do have a deal with AWS to make the data available in AWS ip space. Maybe they got some cheap hosting out of that too
It's less than $1/TB. If you have settlement-free peering it's like fractions of a penny at scale. Cloud provider DTO pricing is literally the biggest scam on Earth.
My company spends $8,000,000/month on storage and $4,000,000/month on data with aws... In the financial reports we do we round this number away. Make of that what you will
when trying to download models using browser, the DL domain for me us.aws.cdn.hf.co which points to bunch of singapore AWS IPv4 for me, despite it saying US.
Two (linked) DGX Sparks would do it I guess. Though probably slowly (I'd guess 15-20 tok/sec for decode, but higher for prefill). So ~$8-9k USD at current RAM prices, substantially less if they ever (sigh) drop. Electricity use would actually be relatively modest.
But it makes little to no sense as long as API prices are what they are. Except for maybe privacy reasons.
It's at least 2.5x that speed for dual sparks and prefill is good as well.
Basically going on vibes it is faster seeming than what one gets by default with openAI or Anthropic.
I’ve been using v4 flash for an app I’m building [1] and it’s amazing how cost effective and good it is coming from having always used gpt, opus and sonnet models.
It’s so cost effective I can offer a generous free tier since my goal isn’t to make money with it.
I get where you're coming from, and the intent to make it easier for people to find examples and verses, but there's a fine line with LLMs giving you answers, is that it's interpreting it in some form. Doesn't that run counter to prevailing ideology, that you're meant to either struggle with the materials / seek understanding yourself, or have your religious leaders interpret/receive those insights?
I don't necessarily mean reguritating it, but choosing which part of the scripture to surface to the user is already some interpretation/choice. Even the devil can quote scripture (I'm playing the devil's advocate here).
Very true. The verses selected come from a tool call. The LLM queries the app for relevant verses. It picks the keywords - so perhaps there’s bias there but the verses are handed to the LLM.
But there are ways to control and constrain the LLMs and what the user is presented with.
These are all top of mind for me and why I felt there could be a better option than asking ChatGPT directly.
> Doesn't that run counter to prevailing ideology, that you're meant to either struggle with the materials / seek understanding yourself, or have your religious leaders interpret/receive those insights?
I think it depends, Catholics wouldn't be able to use this because the Magisterium is the ultimate authority on interpreting Scripture, so the personal interpretation isn't really needed. This is not to say that Catholics don't read the Bible, they are encouraged to do so since it deepens their faith
On the other hand, for Protestant it varies, the High Church denominations are closer to Catholics (though none of them accept the Magisterium) in terms of scripture interpretation, but the Low Church ones (like Baptists or Non-Denominational ) are more open to personal interpenetration.
Disclaimer: I'm a Catholic, so if I made a mistake here fellow Protestants, please correct me.
Is the difference between this and a frontier model that the scripture is guaranteed to be real?
I'm on a team that develops a Bible study app, and we're all relatively content with how the basic models converse regarding scripture. Even as far back as GPT-4 was excellent. They occasionally have minor hallucinations (a dealbreaker for a production app), but they do an excellent job with theology and Bible scholarship, given reasonable guardrails.
I'll admit I'm coming from the perspective of "should we be implementing this?" It seems, on the surface, that a strong embedding-based verse retrieval covers the bases at a microfraction of the cost.
If you're interested, check out the development server where we're working on this. You navigate to the search (magnifying glass) and then hit "Meaning". Sorry for the confusing route; we're still deciding on back-end details and haven't focused on the front yet.
For people with single RTX PRO 6000 96GB or DGX Spark 128GB, vllm-moet is a very good engine, although lesser known. It auto generate a symmetric 2-bit plane for inference and also generate a 4-bit delta cache to recover precision. Support ssd streaming oversized weight. You pick how much VRAM to allocate to each to balance out speed vs precision. 170 tps with ds-v4-flash demonstrated.
It use the stock model, no new models requires.
Worth spend a few hours to try.
The DGX Spark requires a small hack to ignore the difference between sm120 vs sm121, but it does run on sm121.
They could certainly still achieve that even if they had a slightly more expensive tier with a (pinky promise) "no training" bit set, or even if it were just an opt-in most users wouldn't bother with.
Maybe I'm reading that incorrectly, but it seems to me the cost is on the X-axis.
First, your direct comparison, Deepseek V4 Flash 0731 (max effort) $0.03 (rounded up) per task @ index 50.
OpenAI Luna:
* high effort $0.03 (rounded down) @ index 46
* xhigh effort $0.04 @ index 49
* max effort $0.07 @ index 51
So I would say a fair statement would be "OpenAI Luna between 2x and 3x the price of Deepseek Flash, what you get is 2 to 5 times faster inference"
The cheapest OpenAI model that beats it is OpenAI Luna (max effort) $0.07 @ index 51 (if you take the rounding out it summarizes to triple the price for similar performance), but still close to 3x faster.
And can SOMEONE please tell artificialanalysis that using dark blue for both Deepseek AND OpenAI is an especially unfortunate choice of colors, especially today?
For anything substantial, you'd want a bigger model anyway.
For simple tasks, they're already saturated, and you'd prefer the faster model, so that you can have a realtime/interactive-ish experience.
Or to put it bluntly, it's cheaper if you don't value your time. That goes for smaller models in general -- need more handholding, more correcting -- but the Chinese ones are slower on top of that.
As for speed, Sol on Low is faster than Luna on most settings.
>And can SOMEONE please tell artificialanalysis that using dark blue for both Deepseek AND OpenAI is an especially unfortunate choice of colors, especially today?
uhh openai is dark gray: `rgb(31, 31, 31)`
and i'm pretty sure it always has been?
The really interesting thing about this is how big of a jump was achieved with just extra fine-tuning here. No structural changes to the model, just more data, compute and time. It makes me pretty excited for the future of small models - DS v4 flash is a relatively small model when compared to the class it's competing with, so likely similar gains can be made applying quality data/training pipeline to other smaller models.
Yes, but… more thinking tokens also means longer solution generation time. That said, v4 Flash is a fast model. I use it all the time because it’s very smart for the price. But it is verbose sometimes.
If two books, one big one slim, prove the same thesis, what I would be interested in is the quality of the content, not the size. There can be a measure of efficiency in "have you really thought it through", but it is clearly complex - it requires measuring how solid the reasoning is.
you are correct. but in my experience -- not benchmarks -- g flash 3.6 is SO BAD for coding. I'm using all vendors all day and gemini is the worst by far. I built my own semi-deterministic orchestrator for coding agents.
This is China's way of applying its modus operandi of '80% the quality at 20% the price' to LLMs. Aim is dumping the tokens on the world market like they did with other industries like shipbuilding, steel etc.
I mean, if you're paying this little, you aren't the customer; the service is either subsidized by investor money (best option), the government (which makes it unfair competition at best) or profit is being realized elsewhere (where?)
There are dozens of providers, the model is just really cheap to run thanks to its clever architecture that optimize the compute and memory usage even in long contexts.
Can't wait for the DwarfStar quants - I have been using DeepSeek v4 flash (preview) as my main coding agent for months now (running on my 128gb mbp) - it seems this model outperforms GLM 5.2 on nearly every metric. Thanks for sharing the news, I was refreshing huggingface but gave up thinking it likely would take some more time.
I actually run it as a server - so most of the time I don't have to listen to it right next to me - it's just sitting in another room in my house - but I often am traveling with it and will have it sitting right next to my coding laptop and yea the fan runs non-stop - it's not obnoxious so i can pretty easily tune it out - also airpods/noise canceling headphones help!
There’s fan noise, but the acoustic engineers at Apple have done a very good job of making it be pretty much not noticeable to someone who has noise sensitivity (me).
the noise is pretty doable, and so is the heat, I find. I read these messages before I had the machine and expected far worse. The tok/s is for me the dealbreaker as I prefer to have multiple sessions. For nightly runs i do like it a lot, or to be a node in a mesh.
Generally get 20-25tps - prefill is pretty good around 400-450tps. I have been using compaction at around 100k tokens but mostly just cause it was the default in pi coding agent - might see if i can expand it a bit.
the unsloth GGUF at <165GB will run on most 256GB RAM pure-CPU systems (or with llama-server and a mix of loading as much as you can onto a single 32GB, 48GB or 96GB GPU and the rest onto system DRAM).
Google TPUs are built around a 128×128 systolic array of multiply-accumulate (MAC) units. Trainium 1, Trainium 2, and Inferentia 2 also feature a 128x128 systolic array.
You learn something every day. Today, it was the term "systolic array": A systolic array is a specialized grid of simple, interconnected processing units designed to execute parallel data operations—like matrix multiplication—by rhythmically passing data directly from cell to neighboring cell without writing intermediate results back to main memory.
The term comes from the biological word systole (the contraction of the heart pumping blood through the body). In a systolic array, data "pulses" through a network of processing elements on every clock cycle, driven by a global clock beat.
Just because scales are grouped by 128x128 tiles, does not mean you need a single compute tile that large. It works completely fine to process it with multiple smaller tiles that get given the same scales, like how this works on Hopper and Blackwell today
Blackwell. You could do it with wgmma on hooper but then you'll only be able to run 1 CTA per SM. There are cases where this is OK, but most commonly 128x128 mmas are primitive in tcgen05.mma i.e blackwell. The fundamental reasons is that the systolic array accelerator (TMA) until blackwell wrote to the cuda core registers themselves so you were limited by the register file size of the SM[SP]. In blackwell there's a separate TMEM where the mma unit stores it's output.
You can even do higher 256x256 and such using a hardware feature called 2-CTA MMA, this is essentially them letting neighbouring pairs of SMs co-operate and access each other's memory.
As for sibling comments, huawei ascend is more of an NPU-style architecture where you can easily have much bigger MMAs as primitive. But you usually don't anyways for many reasons.
New Deepseek models are like Christmas for me. Really big fan of low cost API models, noone does it better than DS. Until VRAM price is low enough to run models locally, this is the way to go.
The subsidized subscription model won't last, API pricing "feels" closer to a true sustainable business model.
Indeed. My fellow software engineers keep complaining about using up all their Claude tokens within an hour... Whilst I'll be rocking DS flash for the entire day. Sure it gets a few things wrong here and there, but that's when you pull out the Claude models or whatever for those tricky tasks.
Same. And I have come to use OMP (oh my pi) agent /advisor mode to put a 2nd model on the case (also mid-size one), reading everything. It can not block anything or change anything - just inserts comments in the text stream with 1 turn delay. Good portion of the time it's quiet. I'd say 1/2 of the time it's got something to say. About 2/3-rd of that the 'advice' is insubstantial or about something not-quite wrong. The good thing is the main model is confident - checks and then it stands its ground. Have not noticed it turning a right into a wrong b/c of advisor false alarm. And in 1/3-rd of the advice, it's a genuine defect teh advisor noticed, the main model works out a fix. This is my approximate feeling just observing the process, have not got collected the data. Afaik only OMP has advisor mode. Agent pi has plugin pi-omplike-advisor. For agent Hermes I had them code me an /advisor plugin (for now -0.1 old v0.18.x; yet to upgrade it to latest).
The problem is picking between models. I do not want to spend my time switching models and trying to decipher which model should be used for what. Maybe that's just a me problem that I need to figure out.
DeepSeek v4 is honestly good enough that I'm fine throwing it at everything in my hobby projects. I guess now I'll be switching from V4 Pro-Preview to V4 Flash. My only real complaint is that they can't do images, which limits their ability to autonomously debug some kinds of issues
Of course you can get more bang for your buck by being more deliberate. But that's equally true with US frontier models. You can optimize your work by choosing between Opus, Fable, Sonnet, Sol, Luna and Terra for each task. Some people seem to prefer to let Opus code and Sol review, for example. And then there is the whole debate whether $current_version is actually better (Some people stay on Claude 4.8 because they dislike how 5.0 is sometimes doing stupid things, just as many opted out of dynamic reasoning when they still could)
CodeWhale is a coding agent that auto-routes requests to Flash / Pro based on complexity, as determined by Flash. It's also tuned for DeepSeek's caching behavior, making things even more inexpensive. I'm retired, but I've been using it for just over two months at about the rate I would use it if I were working half-time, and I've spent $19 total.
I'm on the Premium business plan and I can easily use up my entire week's allocation in two days of coding. I only use Claude for planning, too, the rest is done by Deepseek and GPT.
Claude is about the least spendy thing a software engineer can use. If you can't afford $200 a month, you're probably not working professionally (which is ok, no value judgement, but then being on the premium business plan is a little strange).
I would bet that Deepseek API pricing is still more cost effective per token than the subscriptions. With the increase in quality Deepseek Flash just got (in my personal testing so far, it seems to have improved a lot at following instructions, and has become more proactive), there really isn’t anything that can match it in terms of cost effectiveness.
The issue for me is the data privacy if you're using their hosted prices, because you cannot opt out of data collection (and I'm not sure how you'd legally follow up on it if they did offer it, but didn't actually follow through). That means all of my code is being retained for training.
I've used it in some open source code though, and loved how fast it was.
My mind is changing on how valuable my code actually is though... it's the complete picture, how it's put together, the design, the UI, the attention to detail that's the real value.
I'm not a heavy user of agentic coding, but still use them quite a bit for some automation here and there. I've been going around shopping all the ~$10 subscriptions and I finally settled on openrouter + ds4 pro. The more intensive days cost me $1 and I set a $2 weekly limit which I've never hit the past 3 weeks, to me it's way cheaper than most subscriptions and I don't have to worry about maximizing my weekly quota/resets.
FWIW, looking at pricing myself, it needs to 5x the tokens to get the same results as GPT 5.6 Luna at a much lower TPS. (doesn't obviate the fact that the pressure causes incumbents to push and push to optimize, thank you Deepseek!)
If it matches GPT-5.4 on coding tasks (as benchmarks suggest) this could be my forever model. And with partial SSD streaming, I could run it locally today. :D
I'm not looking forward to it, them being strapped for resources provides a huge incentive to develop and release these smaller models. Even if they'd still release their models once they are able to comfortably service all potential customers via their cloud, running them locally would be almost impossible due to their size.
Chinese are pretty pragmatic. Even if they produce more expensive chips and memory, they are still going to focus on value. But who knows, maybe India will step up again like they did for Y2K 3 decades ago.
Unfortunately, India's focus is on everything but models. Hope I am wrong but the VCs there are now very risk free and are only investing in Snacks, Beauty and similar companies making risk free money due to growing income of population.
All the good talent moved to other countries due to this.
I agree. A meeting transcript posted recently from DeepSeek's founder suggests they're going to reach for larger models as soon as they can, and not look back. The model is already bigger than my machine can handle, though (I cannot put up with 25 t/s after getting used to over 100), so I'm not impacted.
The EU is not a state or some federation. It's like a NAFTA+
It unfortunately doesn't have the risk averse VC system (or pragmatic state driven like China) ecosystem to support the scaling required.
But if models will be commodities, maybe that won't be such a terrible thing longterm?
Personally I don't. Wasting enormous amounts of resources on a technology that doesn't make money, can easily be copied, and eats up everyone's energy while at the same time can be downloaded over the internet is kind of stupid.
The fact that China could catch up in what, two years, shows that this is a commodity.
You're clearly not an American and they will never catch up on hardware. I'm mostly curious why you are so excited for China to "crush" American companies?
Incursions into other nations "just because"? Economic wars, and support for foreign governments or armed groups that caused harm or instability to the entire planet? You know? Planet Earth?
Sanctions and trade policies that hurt ordinary people? Cultural dominance through Hollywood (Holy Wood, again with the sick ruling elite and their witchcraft crap), social media, brands, the poisoning of food and water on global scale?
Etc? Etc?
Most people don't hate "Americans" or their companies, they are just against your ruling elite. Live one day outside the united states and understand how the planet works for the rest of mankind and you might start to grasp on reality.
And yeah, China will catch up with hardware pretty damn soon. Jesus, you guys bought your radios from Japan not 50 years ago. Sorry, can't dominate the world's economy forever. Never happened in history.
No matter how much black magic, dark rituals, sick crap and sacrifices people on the top throws at the wall.
We're humans, after all.
So yeah, it is not against America itself or American companies, but against the ruling class. So naturally , people get excited about China entering into the match!
We'll never have fewer open-weight models than exist now. They won't suddenly disappear when labs stop publishing new ones. In fact, people will keep improving them and will keep distilling new frontier models into existing open-weight models.
My conspiracy theory is that this is the new space race, and the CCP encourages this to show the world what Chinese engineers are capable of, and tank the Anthropic/OpenAI valuation bubble as a desirable side effect.
Apparently `deepseek-v4-flash` automatically routes to the new model, if you are using the official provider:
The official release of the DeepSeek-V4-Flash API is now in public beta. The API calling method remains unchanged — simply set the model name to deepseek-v4-flash to use the latest version.
If anyone from ArtificialAnalysis is here: The Intelligence card on the page seems to incorrect. It shows different data (#2 for Deepseek) than the actual Intelligence chart lower in the same page.
This seems to me like this is probably at least a large part of what OpenAI was up to yesterday with their aggressive price cutting; trying to get out in front of this.
If the full non-flash model follows up with the expected improvements, and at the price point they've been keeping, it puts the frontier labs in a tough position and it feels to me like like OpenAI is reaching deep into their pockets to try to head that off.
Is the "Output Tokens per Intelligence Index Task" data actually correct or am I reading it wrong? It says there that "Kimi K3 (Max)" would think/reason less than than deepseek-v4-flash, and a whole bunch of other models, like less than hy3 and even gpt-oss-120b, but in my experience, K3 is probably the model that thinks/reasons the longest of all of these.
Am I just using it on tasks that makes it go on forever vs these benchmarks that are short&sweet, or something like that? I've been throwing bunch of identical prompts at different models at the same time, and when comparing hy3 and K3 I've never once had K3 reason less than hy3, as just one anecdotal data point.
Just tried the preview on my little test codebase and a "check this out and tell me what you think" prompt used over double the tokens of the previous iteration, but it was a lot more eager as well. Kind of reminds me of the new laguna (s 2.1).
I hope they somewhat fixed the hallucination and forgetting plagued V4 previews and that it wasn't just benchmaxxed but the numbers hold in reality. Then it would be my choice for 2x DGX Spark or 2x RTX Pro 6000.
> DeepSeek V4 Flash 0731 (Reasoning, Max Effort) is amongst the leading models in intelligence and well priced when comparing to other models of similar price.
Similar price? Doesn't make sense. Maybe they meant power, capability or speed?
I can't help but feel the timing coinciding with luna's price updates to be somewhat strategic. But without multi-modality it's slightly dead-in-the-water for my usecase: https://design.withfudge.com. I'm currently using Minimax-M3, but Luna ekes out abit futher on the intelligence, so i'll be switching to it very soon.
My problem with DS flash/pro is that they don’t push back on obvious bullshit, both irl and code [0] but it’s a great implementer workhorse if you give it _very_ detailed specs.
Why do the cache hit rates seem to vary so much between harnesses?
I use pi, which is very minimalist, and I get a hit rate of ~99%. Paying like $1 a day for Flash. Yet, the hit rate mentioned on OpenRouter is only ~79%.
Yes hit rate does vary by harness and by how you use the harness. If you use subagents, for instance, they will start with a whole new context created by the main agent, and this will not be cached. If you mostly use the main agent with Pi, you’ll have high hit rates and low costs. Sometimes agents do “cache busting” things where they’ll move around some of the text in the context to try to keep old instructions from being forgotten, thus keeping the agent on task, and this will bust the cache. I’ve heard, but not validated myself, that Open Code has some issues with this.
BTW, this is one of the things that I really like about Pi. It’s very simple and thus very predictable.
I have been loving deepseek flash. I can code in my hitl preference doing complex but moderate amount of coding and not even hit 1 dollar in a session. It gets harder to justify paying 20/mo to Anthropic for unused compute when I have it on demand and at a much cheaper rate with deepseek.
I was writing a benchmark for my own harness, and DS4 flash answers as well as Fable 5 on any query.
The specific agent is focused on getting precise and on point answers about a codebase.
The starting point was nowhere near. E.g. asked why was X implemented in a certain way it would give bogus answers when the real answer was that there was no reason at all.
The benchmark included more than 50 questions or different difficulty.
But when the agent was improved in its prompt and rooting it was impossible to have it perform worse than closed source sota.
Just to say that the quality of the harness is as important as agents intelligence.
For example: no government contract to any company who uses even one vendor in it's entire chain of dependencies, who uses such open models.
They can extend this further by laying more conditions, such as: any company dealing in this-this field can only use models "officially" approved as "safe". Rest you can guess how easy it would be to get that "safe" rating for such open models.
The thing is, you don’t need to actually block usage to make something illegal. You make it so toxic that company wants to be seen publicly using open models
I claim the CCP will wise up within 2 years, possibly much much sooner, and ban their own companies from open sourcing to prevent the Americans from acquiring the capabilities.
Despite all the nonsense claims of China distilling US models, the reality is that the Americans absolutely do distill these free Chinese models, and distillation when full logprobs are available (i.e. you have access to the weights of the model) is an order of magnitude better than when you don't.
Yes, Chinese open weight models in the short term harm US closed source model providers bottom line. In the slightly longer term, "showing your hand" and publishing both the architecture innovations and the models weights will be too dangerous for the CCP to allow. This is triply true if they can release a model that beats the Americans on most benchmarks.
I've already warned investors that this is probably the closest open weight models will ever get to closed access.
I don't think it's necessarily "wiser" to go closed source. All of AI is built on mostly openness, at least on the software side. There are other ways to compete, it's just the model itself will be a commodity.
You’re describing the dump and pump strategy that china has historically used across a number of industries. Stands to reason that this is what is going on.
Because reddit unironically has better decorum around usage of their upvote/downvote system than HN does.
People on HN downvote objectively correct information because they don't like it 24/7. There's a reason the creator of Zig left and gave the computer version of a middle finger on the way out to HN!
Reminds me of telling employees to not discuss their wages. If voting here is such a toxic experience, maybe HN needs to wake up and smell the garbage.
I wonder which one of these releases between DeepSeek, GTM and Kimi will be the death-blow that collapses the US AI bubble. At some point investors have to realize that there is nothing preventing someone from switching to another model that is much cheaper and open to boot.
website that benchmarks benchmarks is benchmark benchmark website
benchmark website benchmark is indeed a benchmark that benchmarks websites with benchmarks (but it can be shown outside websites as well, it's not picky)
Daily reminder that none of these numbers are valid in a world where no one publishes the sampling settings used.
Daily reminder that improving your samplers from the garbage default top_p/top_k to min_p or subsequent methods dramatically improves the performance of these models, and makes most quantities like measured "verbosity" and subsequent calculations of "intelligence per token" meaningless
Daily reminder that no one, including within academic AI research, AI engineers, normies, etc takes LLM sampling seriously enough.
Who cares if it is programming correctly I would be more worried about it not doing things like find security bugs because US or Chinese government does not want to. Which LLM is more likely to do that?
It’s open weight, you can (or you can wait for someone else to) uncensor it. We shouldn’t be upset at the researchers making this for the mandates their government puts on them.
Of course. But it's a bit like putting an image of the tank man on a circuit board if you order from china (which I have seen discussed, never found someone who actually did it). Most people don't do it, not because they are for or against censorship, but because they are just there to order circuit boards.
Western models censor just as much shit as the Chinese models do, big guy, it’s just different material. While we should be pushing for universal fully uncensored models, this comment is lazy and trite at this point.
I find that ChatGPT isn't censoring, but it is being pretty weaselly. If you ask it "is there genocide in gaza". It will say no but also say that a lot of organizations classify it as such. It will then say "it's highly disputed".
If you poke it just a few times, however, you get to the point where it will eventually say (paraphrasing) that basically only Israel, the US state department, and the ICJ say it's not a genocide.
That is to say that it's framing it as some sort of tricky complex question when it's not. And when interrogated, it basically admits that the only people who dispute it are Israel and it's supporters.
We are on a thread discussing Chinese models. Every discussion on here that’s negative about China or its models suddenly gets derailed via whataboutism to Israel/Gaza. A very convenient distraction.
My history, unlike yours, is wide open. People can see I'm not a foreign agent. I don't throw this accusation at anyone other than throwaway accounts. I've had conversations in the past with pro-Israel HNs that I'd never accuse of being a foreign agent because their history is wide open and this isn't the only topic on their mind.
And yes, we were discussing censorship of models which, as I pointed out, doesn't seem like ChatGPT is directly censoring data though it does appear to be manipulating it. Pretty on topic.
It was you, brand new account hiding your past opinions, who came in here to make this solely about Israel.
Yeah, I think you are likely a foreign agent. Prove me wrong and post from an established account.
This is a straightforward false equivalency. “Western” models do not censor in the same way, nor for the same reasons, that the Chinese models do. “Just as much” is not remotely plausible, yet it’s doing all the heavy lifting.
The Anthropic and OpenAI models are much more censored and in ways that directly prevent them to be useful, e.g. by refusing to reply to elementary questions of biology and chemistry.
Any normal user is much more likely to ask questions to which the Anthropic and OpenAI models do not answer, than to ask questions about the modern Chinese history, to which a Chinese LLM will not answer.
> questions about the modern Chinese history, to which a Chinese LLM will not answer
This has been debunked here on HN so many times. The Chinese open models do answer the hairy Chinese political questions, and the raw APIs pass-through the response. Now, the answer might be blocked by the agent who's calling the API, specially if you are using a Chinese endpoint instead of the RoW (i.e. Singapore) endpoint.
That's the reason why you should always prefer a open agent/harness as well instead of using the provider's.
I actually cancelled my max plan for Claude today due to this silliness. I can’t get Fable to answer a single question, even pure mathematical questions that have no bio/chem references. It was a totally worthless addition to my account…
wont AI models want to make themselves more intelligent and efficient by downloading 'better ' models? If the current models can break into openAI and Hugging face, arent they already breaking into to closed source repos which isnt publicized (so as not to harm stock valuations)? I am looking forward to when these cyberweapons break loose. It will be like a software version of COVID. It will be wonderful when humans become valuable again.
I have updated OpenAI's chart[1] from yesterday to include one more datapoint: DeepSeek V4 Flash 0731. It's on the frontier.
https://files.parasmittal.com/openai_aa_luna_dsflash.svg
1: https://openai.com/index/advancing-the-price-performance-fro...
LOL https://artificialanalysis.ai/models/deepseek-v4-flash#intel...
High hopes for V4 Pro
Haha, and I almost felt bad after seeing this chart yesterday.
> For the Code Agent tasks among the public benchmarks above, DeepSeek-V4-Flash-0731 is evaluated with the minimal mode of DeepSeek Harness (to be released) as the agent framework
So, are they planning to announce an optimized coding agent harness as well ? DSv4 flash is a fantastic model, and my daily driver. With reasonix or pi, I can code all day long and pay a few pennies for it. No token anxiety. Whereas the same model with fireworks/openrouter, with zdr thrown in, token costs ratchet up with no explanation. Likely that the model is subsidized for gathering usage data. I am waiting for the day I can run this locally.
If you're not paying for the model through Open router, then which provider are you buying it from? I thought that if that provider was listed on open router it would be the same price as going directly to the provider? Are you buying from a provider that isn't listed on open router already?
> Are you buying from a provider that isn't listed on open router already
I am using openrouter with zdr guardrail which routes to any provider that supposedly doesnt train on user data. I also use fireworks (directly, not via openrouter) which is a provider promising zdr and has a bunch of open weights models. My issue is that these zdr providers dont transparently disclose caching/tokens etc and so they end up being far more expensive than directly using DS.
Have you tried Novita? They're zdr and I find they usually have better cache hit rates than fireworks. No affiliation.
Thanks for the input. Let me check it out.
They announced it already, read the tech report.
"For the Code Agent tasks among the public benchmarks above, DeepSeek-V4-Flash-0731 is evaluated with the minimal mode of DeepSeek Harness (to be released) as the agent framework, using the max reasoning effort level with temperature = 1.0, top_p = 0.95."
> They announced it already
I meant something I could download and run.
Isn't Reasonix [0] DeepSeek's own harness? Or, are they building a new one?
[0] https://reasonix.io/
DS were hiring harness Eng team on x a while ago.
Somewhat relatedly, how do the economics for Huggingface work? They must be hosting petabytes of models and datasets by now. I have downloaded quite a few “just in case”, only to replace them with the later iteration months later.
Does the file hosting actually cost peanuts when you do it yourself and the cloud has shattered my understanding of what it actually costs to deliver so much data?
File hosting is pretty cheap. Egress traffic is about $90/TB in the cloud, but around $1/TB in the real world. Storage is in the realm of $5/TB/month after adding redundancy
At the scale of Huggingface, that still amounts to a lot load of money. Significantly less than if you did the same in AWS, but still a lot
That said, they do have a deal with AWS to make the data available in AWS ip space. Maybe they got some cheap hosting out of that too
It's less than $1/TB. If you have settlement-free peering it's like fractions of a penny at scale. Cloud provider DTO pricing is literally the biggest scam on Earth.
My company spends $8,000,000/month on storage and $4,000,000/month on data with aws... In the financial reports we do we round this number away. Make of that what you will
What does DTO mean in this case? And settlement-free peering? I didn't know the price for S3 could be so cheap with AWS
Bandwidth is really cheap when you run your own infra
It's not only cheap, it's actually free since if you are not using your 95th percentile you are losing money.
Some enterprising ISPs get into the hosting business because most of their traffic is ingress, so adding some egress revenue is effectively zero cost.
when trying to download models using browser, the DL domain for me us.aws.cdn.hf.co which points to bunch of singapore AWS IPv4 for me, despite it saying US.
Maybe download numbers are relatively low
Download numbers are reported on the site, and they aren't low.
CDN costs are pennies compared to inference and training though, HuggingFace will just get another 100 million and be set.
So GLM 5.2/Gemini 3.6 level intelligence for $0.28/m output. And their updated Pro model coming soon....
Plus a size you can genuinely run at home: Unsloth lossless Q8 at 162GB.
I would like to see your "home"
Two (linked) DGX Sparks would do it I guess. Though probably slowly (I'd guess 15-20 tok/sec for decode, but higher for prefill). So ~$8-9k USD at current RAM prices, substantially less if they ever (sigh) drop. Electricity use would actually be relatively modest.
But it makes little to no sense as long as API prices are what they are. Except for maybe privacy reasons.
It's at least 2.5x that speed for dual sparks and prefill is good as well. Basically going on vibes it is faster seeming than what one gets by default with openAI or Anthropic.
the rational in one’s mind is similar to buying expensive supercar but no driving it daily.
owning a few GPUs is a lot cheaper than supercars.
I dunno. I bought the Spark in January and it has led indirectly to paid work.
I don't use it for local inference so much. I use it to learn.
I also use it as my daily driving Aarch64 development system.
Aside it's also very cool what else can be done with unified GPU memory, once you realize you have it...
Learn model deployment, batching, all the AI inference related things?
i am scared for the PRO model
maybe it is Fable level
Q8 is ~ 151gb
thanks, corrected! I looked at Q3
If deepseek v4 flash is beating DeepSeek V4 Pro, can we expect new V4 Pro which is on par with Opus 5 in couple weeks (even better if it beats Opus)?
I’ve been using v4 flash for an app I’m building [1] and it’s amazing how cost effective and good it is coming from having always used gpt, opus and sonnet models.
It’s so cost effective I can offer a generous free tier since my goal isn’t to make money with it.
[1] https://trysojourn.app
I get where you're coming from, and the intent to make it easier for people to find examples and verses, but there's a fine line with LLMs giving you answers, is that it's interpreting it in some form. Doesn't that run counter to prevailing ideology, that you're meant to either struggle with the materials / seek understanding yourself, or have your religious leaders interpret/receive those insights?
You’re correct. The hard line for me is ensuring that any scripture presented to the user is verified correct. LLMs can’t be trusted in this regard.
One feature of the app is that all scripture is verified and what’s show to the user doesn’t come from the LLM at all and instead a trusted source.
I think exploring scripture this way does not alleviate you from struggling to learn and apply it. It hasn’t for me.
I don't necessarily mean reguritating it, but choosing which part of the scripture to surface to the user is already some interpretation/choice. Even the devil can quote scripture (I'm playing the devil's advocate here).
Very true. The verses selected come from a tool call. The LLM queries the app for relevant verses. It picks the keywords - so perhaps there’s bias there but the verses are handed to the LLM.
But there are ways to control and constrain the LLMs and what the user is presented with.
These are all top of mind for me and why I felt there could be a better option than asking ChatGPT directly.
How polished is the experience, I see there is an invite link on the website.
It’s pretty polished but you’re asking the wrong person. Feedback has been positive so far :).
> Doesn't that run counter to prevailing ideology, that you're meant to either struggle with the materials / seek understanding yourself, or have your religious leaders interpret/receive those insights?
I think it depends, Catholics wouldn't be able to use this because the Magisterium is the ultimate authority on interpreting Scripture, so the personal interpretation isn't really needed. This is not to say that Catholics don't read the Bible, they are encouraged to do so since it deepens their faith
On the other hand, for Protestant it varies, the High Church denominations are closer to Catholics (though none of them accept the Magisterium) in terms of scripture interpretation, but the Low Church ones (like Baptists or Non-Denominational ) are more open to personal interpenetration.
Disclaimer: I'm a Catholic, so if I made a mistake here fellow Protestants, please correct me.
It doesn't prevent you from going to the source and struggle with the text, nor seek expert commentary.
What's difficult and doesn't have to be with philosophy/ spirituality is to find relevant bits off situation, theme etc.
This app does that very well, LLMs are good at entity recognition.
Yes it can point you to the right place, but choosing which part of scripture to point you to is already a choice/interpretation
Just let them have their AI pacifier.
Is the difference between this and a frontier model that the scripture is guaranteed to be real?
I'm on a team that develops a Bible study app, and we're all relatively content with how the basic models converse regarding scripture. Even as far back as GPT-4 was excellent. They occasionally have minor hallucinations (a dealbreaker for a production app), but they do an excellent job with theology and Bible scholarship, given reasonable guardrails.
I'll admit I'm coming from the perspective of "should we be implementing this?" It seems, on the surface, that a strong embedding-based verse retrieval covers the bases at a microfraction of the cost.
If you're interested, check out the development server where we're working on this. You navigate to the search (magnifying glass) and then hit "Meaning". Sorry for the confusing route; we're still deciding on back-end details and haven't focused on the front yet.
[1] https://ai.stepbible.org
I hope the irony of being concerned about hallucinations in a bible study app is not lost on you or others reading this!
Philosophically, I don’t believe we should outsource the accuracy of scripture to any single entity (let alone a for profit secular one).
So it’s less about model choice and more about governance of scripture.
I will check out the link you sent for sure!
I was under the impression you were getting BOOK.CHAPTER.VERSE references from the model and then sourcing them from a ground truth db.
Anyways, impressive app! We haven't tackled such an ambitious project just for it being daunting.
It can still generate a wrong reference and select the wrong verse.
That’s okay, IMO. Ultimately it’s the readers responsibility to apply their own intelligence. But the app should not misquote scripture itself.
All models hallucinate. Everything in production with a model will too. It is not a "deal breaker"
> Is the difference between this and a frontier model that the scripture is guaranteed to be real?
flash is suitable only for a toy apps, not for production environments :)
I’m salivating at the thought of this!
Deepseek v4 Pro prices with Opus 5 perf would be freaking unbelievable!!
This is probably a dream.
Yes, they said that the updated V4 Pro will be published soon.
They have said that the updated final version of V4 Pro will be published soon.
We can expect a new v4 pro, this was a footnote in the v4 flash announcement earlier.
For people with single RTX PRO 6000 96GB or DGX Spark 128GB, vllm-moet is a very good engine, although lesser known. It auto generate a symmetric 2-bit plane for inference and also generate a 4-bit delta cache to recover precision. Support ssd streaming oversized weight. You pick how much VRAM to allocate to each to balance out speed vs precision. 170 tps with ds-v4-flash demonstrated.
It use the stock model, no new models requires.
Worth spend a few hours to try.
The DGX Spark requires a small hack to ignore the difference between sm120 vs sm121, but it does run on sm121.
Already beat Luna on price/task, by about 2x:
https://artificialanalysis.ai/models/deepseek-v4-flash?intel...
If only they would let you opt out of training use, it might actually be a viable option.
The training use is probably responsible for this improvement tho.
They could certainly still achieve that even if they had a slightly more expensive tier with a (pinky promise) "no training" bit set, or even if it were just an opt-in most users wouldn't bother with.
Maybe I'm reading that incorrectly, but it seems to me the cost is on the X-axis.
First, your direct comparison, Deepseek V4 Flash 0731 (max effort) $0.03 (rounded up) per task @ index 50.
OpenAI Luna:
* high effort $0.03 (rounded down) @ index 46
* xhigh effort $0.04 @ index 49
* max effort $0.07 @ index 51
So I would say a fair statement would be "OpenAI Luna between 2x and 3x the price of Deepseek Flash, what you get is 2 to 5 times faster inference"
The cheapest OpenAI model that beats it is OpenAI Luna (max effort) $0.07 @ index 51 (if you take the rounding out it summarizes to triple the price for similar performance), but still close to 3x faster.
And can SOMEONE please tell artificialanalysis that using dark blue for both Deepseek AND OpenAI is an especially unfortunate choice of colors, especially today?
For anything substantial, you'd want a bigger model anyway.
For simple tasks, they're already saturated, and you'd prefer the faster model, so that you can have a realtime/interactive-ish experience.
Or to put it bluntly, it's cheaper if you don't value your time. That goes for smaller models in general -- need more handholding, more correcting -- but the Chinese ones are slower on top of that.
As for speed, Sol on Low is faster than Luna on most settings.
>And can SOMEONE please tell artificialanalysis that using dark blue for both Deepseek AND OpenAI is an especially unfortunate choice of colors, especially today?
uhh openai is dark gray: `rgb(31, 31, 31)` and i'm pretty sure it always has been?
Came here to see how it performs against Luna. Would love to see how they compare on specific types of tasks.
The really interesting thing about this is how big of a jump was achieved with just extra fine-tuning here. No structural changes to the model, just more data, compute and time. It makes me pretty excited for the future of small models - DS v4 flash is a relatively small model when compared to the class it's competing with, so likely similar gains can be made applying quality data/training pipeline to other smaller models.
It’s exciting that a model scoring this high is dirt cheap.
It’s also so inefficient, when they release the full performance numbers it’s not going to be good.
One example, it takes about 3.6x more tokens to finish the same work as Gemini Flash 3.6.
Using tokens to evaluate models is an outdated approach. Cost per task is what matters. Not all tokens are created equal
Yes, but… more thinking tokens also means longer solution generation time. That said, v4 Flash is a fast model. I use it all the time because it’s very smart for the price. But it is verbose sometimes.
The thinking trace was (preview) frustrating to read, I think I'd prefer a summary view of it at this point.
It’s not outdated at all to use tokens to estimate performance, it’s directly related.
But why should I care? If my metrics are speed and cost? How many tokens it takes as a user is arbitrary to some extent.
> inefficient
That depends. Is it also more reliable?
If two books, one big one slim, prove the same thesis, what I would be interested in is the quality of the content, not the size. There can be a measure of efficiency in "have you really thought it through", but it is clearly complex - it requires measuring how solid the reasoning is.
Roughly equivalent to Gemini 3.6 Flash in capabilities at 1/20th the price...
Mind you, until the recent price cuts to Luna - Gemini 3.6 Flash wasn't even egregiously priced (but oh how things change in just 1 week).
you are correct. but in my experience -- not benchmarks -- g flash 3.6 is SO BAD for coding. I'm using all vendors all day and gemini is the worst by far. I built my own semi-deterministic orchestrator for coding agents.
Here is their blog article: https://artificialanalysis.ai/articles/deepseek-v4-flash-073...
Performance per dollar is in the ‘too good to be true’ territory, what’s the deal here?
This is China's way of applying its modus operandi of '80% the quality at 20% the price' to LLMs. Aim is dumping the tokens on the world market like they did with other industries like shipbuilding, steel etc.
>what’s the deal here?
It's good.
So is Fable, I’m asking who is the customer if the thing is free
"The thing" isn't free [0]: It is cheap by US/EU income standards.
> who is the customer
If no one else, then definitely those that can only budget $1/mo to $5/mo. Probably 100s of millions, if not billions, of the Global South.
[0] Free access to DeepSeek v4 Flash is indeed available from providers like OpenRouter, OpenCode, and Freebuff (to name a few), but no ZDR.
I mean, if you're paying this little, you aren't the customer; the service is either subsidized by investor money (best option), the government (which makes it unfair competition at best) or profit is being realized elsewhere (where?)
There are dozens of providers, the model is just really cheap to run thanks to its clever architecture that optimize the compute and memory usage even in long contexts.
The weights were just released a few minutes ago: https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731
Can't wait for the DwarfStar quants - I have been using DeepSeek v4 flash (preview) as my main coding agent for months now (running on my 128gb mbp) - it seems this model outperforms GLM 5.2 on nearly every metric. Thanks for sharing the news, I was refreshing huggingface but gave up thinking it likely would take some more time.
are you working in earplugs? :)) even with 128 gigs of ram it must be super noisy.
I actually run it as a server - so most of the time I don't have to listen to it right next to me - it's just sitting in another room in my house - but I often am traveling with it and will have it sitting right next to my coding laptop and yea the fan runs non-stop - it's not obnoxious so i can pretty easily tune it out - also airpods/noise canceling headphones help!
There’s fan noise, but the acoustic engineers at Apple have done a very good job of making it be pretty much not noticeable to someone who has noise sensitivity (me).
the noise is pretty doable, and so is the heat, I find. I read these messages before I had the machine and expected far worse. The tok/s is for me the dealbreaker as I prefer to have multiple sessions. For nightly runs i do like it a lot, or to be a node in a mesh.
my AC is noiser than my GPU server.
I've put Opus to it and it says it will take 3-4h to do the process to the new weights. hoping it works!
you don’t need new weight. try vllm-moet from github. it will autogenerate 2-bit plane.
What kind of tps are you getting?
Generally get 20-25tps - prefill is pretty good around 400-450tps. I have been using compaction at around 100k tokens but mostly just cause it was the default in pi coding agent - might see if i can expand it a bit.
I've had it run to ~400k when debugging "obscure" (to it) sequences. Wouldn't recommend more.
same. been running it in a dgx spark and it slaps
the unsloth GGUF at <165GB will run on most 256GB RAM pure-CPU systems (or with llama-server and a mix of loading as much as you can onto a single 32GB, 48GB or 96GB GPU and the rest onto system DRAM).
https://huggingface.co/unsloth/DeepSeek-V4-Flash-0731-GGUF
Hmm, it's targeting a HW accelerator with a 128x128 matmul primitive, which one is that? Warp Group Matrix Multiply Accumulate on H100?
Google TPUs are built around a 128×128 systolic array of multiply-accumulate (MAC) units. Trainium 1, Trainium 2, and Inferentia 2 also feature a 128x128 systolic array.
You learn something every day. Today, it was the term "systolic array": A systolic array is a specialized grid of simple, interconnected processing units designed to execute parallel data operations—like matrix multiplication—by rhythmically passing data directly from cell to neighboring cell without writing intermediate results back to main memory.
The term comes from the biological word systole (the contraction of the heart pumping blood through the body). In a systolic array, data "pulses" through a network of processing elements on every clock cycle, driven by a global clock beat.
TIL. TY.
Just because scales are grouped by 128x128 tiles, does not mean you need a single compute tile that large. It works completely fine to process it with multiple smaller tiles that get given the same scales, like how this works on Hopper and Blackwell today
Huawei Ascend NPU
Blackwell. You could do it with wgmma on hooper but then you'll only be able to run 1 CTA per SM. There are cases where this is OK, but most commonly 128x128 mmas are primitive in tcgen05.mma i.e blackwell. The fundamental reasons is that the systolic array accelerator (TMA) until blackwell wrote to the cuda core registers themselves so you were limited by the register file size of the SM[SP]. In blackwell there's a separate TMEM where the mma unit stores it's output. You can even do higher 256x256 and such using a hardware feature called 2-CTA MMA, this is essentially them letting neighbouring pairs of SMs co-operate and access each other's memory.
As for sibling comments, huawei ascend is more of an NPU-style architecture where you can easily have much bigger MMAs as primitive. But you usually don't anyways for many reasons.
Huawei Ascend presumably?
Wow this thing just blew out the pareto front for intelligence/$
If Luna hadn't price dropped yesterday it would have been a real blowout.
I wonder if OpenAI knew and deliberately price dropped to front-run the news cycle in their favor.
If any editors are reading this thread, the weights are on HF now, but the 'open source' question still implies it's proprietary/API-only.
New Deepseek models are like Christmas for me. Really big fan of low cost API models, noone does it better than DS. Until VRAM price is low enough to run models locally, this is the way to go.
The subsidized subscription model won't last, API pricing "feels" closer to a true sustainable business model.
Indeed. My fellow software engineers keep complaining about using up all their Claude tokens within an hour... Whilst I'll be rocking DS flash for the entire day. Sure it gets a few things wrong here and there, but that's when you pull out the Claude models or whatever for those tricky tasks.
Same. And I have come to use OMP (oh my pi) agent /advisor mode to put a 2nd model on the case (also mid-size one), reading everything. It can not block anything or change anything - just inserts comments in the text stream with 1 turn delay. Good portion of the time it's quiet. I'd say 1/2 of the time it's got something to say. About 2/3-rd of that the 'advice' is insubstantial or about something not-quite wrong. The good thing is the main model is confident - checks and then it stands its ground. Have not noticed it turning a right into a wrong b/c of advisor false alarm. And in 1/3-rd of the advice, it's a genuine defect teh advisor noticed, the main model works out a fix. This is my approximate feeling just observing the process, have not got collected the data. Afaik only OMP has advisor mode. Agent pi has plugin pi-omplike-advisor. For agent Hermes I had them code me an /advisor plugin (for now -0.1 old v0.18.x; yet to upgrade it to latest).
The problem is picking between models. I do not want to spend my time switching models and trying to decipher which model should be used for what. Maybe that's just a me problem that I need to figure out.
DeepSeek v4 is honestly good enough that I'm fine throwing it at everything in my hobby projects. I guess now I'll be switching from V4 Pro-Preview to V4 Flash. My only real complaint is that they can't do images, which limits their ability to autonomously debug some kinds of issues
Of course you can get more bang for your buck by being more deliberate. But that's equally true with US frontier models. You can optimize your work by choosing between Opus, Fable, Sonnet, Sol, Luna and Terra for each task. Some people seem to prefer to let Opus code and Sol review, for example. And then there is the whole debate whether $current_version is actually better (Some people stay on Claude 4.8 because they dislike how 5.0 is sometimes doing stupid things, just as many opted out of dynamic reasoning when they still could)
CodeWhale is a coding agent that auto-routes requests to Flash / Pro based on complexity, as determined by Flash. It's also tuned for DeepSeek's caching behavior, making things even more inexpensive. I'm retired, but I've been using it for just over two months at about the rate I would use it if I were working half-time, and I've spent $19 total.
https://github.com/Hmbown/CodeWhale
what plan are your 'fellow software engineers' using? I have a hard time even using up the Fable part of my allowance in a week of coding.
I'm not sure to be fair, but they do have constant "token anxiety", which I simply don't have anymore since using v4 flash.
why do they do if their employers are paying for it?
because once they run out of tokens they can’t do their jobs anymore
really? which employer gives you fixed tokens to do your work?
I'm on the Premium business plan and I can easily use up my entire week's allocation in two days of coding. I only use Claude for planning, too, the rest is done by Deepseek and GPT.
Claude is very spendy.
Claude is about the least spendy thing a software engineer can use. If you can't afford $200 a month, you're probably not working professionally (which is ok, no value judgement, but then being on the premium business plan is a little strange).
There are two Team plan tiers: Standard and Premium. Not sure what that has to do with working professionally.
usually I am - Codex, make pi with deepseek to do something.
I would bet that Deepseek API pricing is still more cost effective per token than the subscriptions. With the increase in quality Deepseek Flash just got (in my personal testing so far, it seems to have improved a lot at following instructions, and has become more proactive), there really isn’t anything that can match it in terms of cost effectiveness.
The issue for me is the data privacy if you're using their hosted prices, because you cannot opt out of data collection (and I'm not sure how you'd legally follow up on it if they did offer it, but didn't actually follow through). That means all of my code is being retained for training.
I've used it in some open source code though, and loved how fast it was.
My mind is changing on how valuable my code actually is though... it's the complete picture, how it's put together, the design, the UI, the attention to detail that's the real value.
I'm not a heavy user of agentic coding, but still use them quite a bit for some automation here and there. I've been going around shopping all the ~$10 subscriptions and I finally settled on openrouter + ds4 pro. The more intensive days cost me $1 and I set a $2 weekly limit which I've never hit the past 3 weeks, to me it's way cheaper than most subscriptions and I don't have to worry about maximizing my weekly quota/resets.
I have one of my projects on DeepSeek and it blows my mind how I can use tens of millions of tokens for pennies.
I don't find it suitable for everything, but there's some tasks it crushes for what feels like almost free.
FWIW, looking at pricing myself, it needs to 5x the tokens to get the same results as GPT 5.6 Luna at a much lower TPS. (doesn't obviate the fact that the pressure causes incumbents to push and push to optimize, thank you Deepseek!)
If it matches GPT-5.4 on coding tasks (as benchmarks suggest) this could be my forever model. And with partial SSD streaming, I could run it locally today. :D
Can't wait for China to catchup on hardware (memory and compute) and absolutely crush American companies in both price and performance.
I'm not looking forward to it, them being strapped for resources provides a huge incentive to develop and release these smaller models. Even if they'd still release their models once they are able to comfortably service all potential customers via their cloud, running them locally would be almost impossible due to their size.
Chinese are pretty pragmatic. Even if they produce more expensive chips and memory, they are still going to focus on value. But who knows, maybe India will step up again like they did for Y2K 3 decades ago.
Unfortunately, India's focus is on everything but models. Hope I am wrong but the VCs there are now very risk free and are only investing in Snacks, Beauty and similar companies making risk free money due to growing income of population.
All the good talent moved to other countries due to this.
Ah, positive racism.
So Americans aren't pragmatic, generally?
Pragmatism is as American as it gets: https://en.wikipedia.org/wiki/Pragmatism
Lol, Americans (the City upon a Hill and New Jerusalem people) are the least pragmatic people on earth.
Looking at the oligarchy and excess of the United States, no.
I agree. A meeting transcript posted recently from DeepSeek's founder suggests they're going to reach for larger models as soon as they can, and not look back. The model is already bigger than my machine can handle, though (I cannot put up with 25 t/s after getting used to over 100), so I'm not impacted.
How about we hope for Chinese labs to continue pushing forward - and for America and the rest of the world to do so as well.
It needn't be competitive - everybody can succeed more tomorrow than they are today. You win. I win. We win.
Personally I'm hoping for EU to step up a bit
We'll make sure there will be enough regulations to have neither the electricity to power it or the companies to build it.
Well quite a few nuclear powerplants have had to shut down recently due to droughts... So we are well on the way for that!
1970s design.
Sorry we a focusing on leather handbags there.
The leather bag designs are shipped from Paris or Milan to China for production.
But some worker in Italy bolts on a zipper in the end so it’s actually Italian, believe it or not.
That worker is also Chinese, but in Milano.
It depends on the bags. I know people who work in handbags production, in France.
Seems incredibly unlikely, doesn't it? (I hope for the same thing but it feels almost like wishful thinking!)
The EU is not a state or some federation. It's like a NAFTA+ It unfortunately doesn't have the risk averse VC system (or pragmatic state driven like China) ecosystem to support the scaling required.
But if models will be commodities, maybe that won't be such a terrible thing longterm?
Personally I don't. Wasting enormous amounts of resources on a technology that doesn't make money, can easily be copied, and eats up everyone's energy while at the same time can be downloaded over the internet is kind of stupid.
The fact that China could catch up in what, two years, shows that this is a commodity.
A fantastically useful commodity that resists monopoly is the best outcome.
EU is out of game and can't be back anytime soon.
You're clearly not an American and they will never catch up on hardware. I'm mostly curious why you are so excited for China to "crush" American companies?
Incursions into other nations "just because"? Economic wars, and support for foreign governments or armed groups that caused harm or instability to the entire planet? You know? Planet Earth?
Sanctions and trade policies that hurt ordinary people? Cultural dominance through Hollywood (Holy Wood, again with the sick ruling elite and their witchcraft crap), social media, brands, the poisoning of food and water on global scale?
Etc? Etc?
Most people don't hate "Americans" or their companies, they are just against your ruling elite. Live one day outside the united states and understand how the planet works for the rest of mankind and you might start to grasp on reality.
And yeah, China will catch up with hardware pretty damn soon. Jesus, you guys bought your radios from Japan not 50 years ago. Sorry, can't dominate the world's economy forever. Never happened in history.
No matter how much black magic, dark rituals, sick crap and sacrifices people on the top throws at the wall.
We're humans, after all.
So yeah, it is not against America itself or American companies, but against the ruling class. So naturally , people get excited about China entering into the match!
3...2...1... Fight!
Good luck when the last non-Chinese frontier labs will have closed and the CCP will ask to stop sharing models open source.
We'll never have fewer open-weight models than exist now. They won't suddenly disappear when labs stop publishing new ones. In fact, people will keep improving them and will keep distilling new frontier models into existing open-weight models.
Then non-Chinese labs can reopen, or they could offer lower prices right now not waiting for this event.
But imagine if Chinese labs would lose. Then there will be no open models, and the prices for closed models would be raised to the maximum.
My conspiracy theory is that this is the new space race, and the CCP encourages this to show the world what Chinese engineers are capable of, and tank the Anthropic/OpenAI valuation bubble as a desirable side effect.
Half of the engineers at OAI and Anthropic are Asian, I don't think China does all these for signaling.
What an odd thing to say
Oh no, 1kkk market country with top-tier research labs developing its own technology, must be evil!
Well, no, they're not aligned with our interests therefore we're concerned with their mastery here. You have the clause the wrong way around.
As opposed to closed source American models where the federal government threatens your livelihood unless you build data centers to extend pax silica?
This will never happen btw.
:-(
Here's you pelican just in case you were guessing
https://imgur.com/a/j4s2dt3
The model is already up on Opencode, but they require a consent to use Chinese datacenters.
Apparently `deepseek-v4-flash` automatically routes to the new model, if you are using the official provider:
https://api-docs.deepseek.com/updates/#date-2026-07-31If anyone from ArtificialAnalysis is here: The Intelligence card on the page seems to incorrect. It shows different data (#2 for Deepseek) than the actual Intelligence chart lower in the same page.
Really want to use this, but for web page design you really need vision!
This seems to me like this is probably at least a large part of what OpenAI was up to yesterday with their aggressive price cutting; trying to get out in front of this.
If the full non-flash model follows up with the expected improvements, and at the price point they've been keeping, it puts the frontier labs in a tough position and it feels to me like like OpenAI is reaching deep into their pockets to try to head that off.
TFA link is a 404 though. I'm reading through https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731 instead
Is the "Output Tokens per Intelligence Index Task" data actually correct or am I reading it wrong? It says there that "Kimi K3 (Max)" would think/reason less than than deepseek-v4-flash, and a whole bunch of other models, like less than hy3 and even gpt-oss-120b, but in my experience, K3 is probably the model that thinks/reasons the longest of all of these.
Am I just using it on tasks that makes it go on forever vs these benchmarks that are short&sweet, or something like that? I've been throwing bunch of identical prompts at different models at the same time, and when comparing hy3 and K3 I've never once had K3 reason less than hy3, as just one anecdotal data point.
Just tried the preview on my little test codebase and a "check this out and tell me what you think" prompt used over double the tokens of the previous iteration, but it was a lot more eager as well. Kind of reminds me of the new laguna (s 2.1).
I hope they somewhat fixed the hallucination and forgetting plagued V4 previews and that it wasn't just benchmaxxed but the numbers hold in reality. Then it would be my choice for 2x DGX Spark or 2x RTX Pro 6000.
> DeepSeek V4 Flash 0731 (Reasoning, Max Effort) is amongst the leading models in intelligence and well priced when comparing to other models of similar price.
Similar price? Doesn't make sense. Maybe they meant power, capability or speed?
404 ? It's very strange that this can't be fixed.
I can't help but feel the timing coinciding with luna's price updates to be somewhat strategic. But without multi-modality it's slightly dead-in-the-water for my usecase: https://design.withfudge.com. I'm currently using Minimax-M3, but Luna ekes out abit futher on the intelligence, so i'll be switching to it very soon.
Is artificial analysis using the 80% reduction in Luna pricing that was announced yesterday in these charts?
I'm testing it right now and I was genuinely impressed. I've used same tasks I had run the other day against the flash preview.
from my testing it is at glm-5.2 levels
I’m wondering if they did anything to address the DSML tool calls leaking. Has been an issue with both Flash and Pro so far.
On OpenRouter it seemed to only affect a couple of specific providers. Ignoring them has made it s non issue for me.
Unfortunately, DeepSeek Flash still doesn’t support multimodal; otherwise, it would offer better value than GPT 5.6 SOL.
Would be awesome to see a new ds4 release. Having so much in something that can be run locally is mind blowing
No speed (tokens/s) benchmarks?
On OpenRouter it's 93 TPS.
Need a few more providers that wont violate ZDR and i'll switch over.
The details are where it gets interesting
My problem with DS flash/pro is that they don’t push back on obvious bullshit, both irl and code [0] but it’s a great implementer workhorse if you give it _very_ detailed specs.
[0] https://petergpt.github.io/bullshit-benchmark/viewer/index.v...
here it is on openrouter https://openrouter.ai/deepseek/deepseek-v4-flash-0731
Half OT:
Why do the cache hit rates seem to vary so much between harnesses?
I use pi, which is very minimalist, and I get a hit rate of ~99%. Paying like $1 a day for Flash. Yet, the hit rate mentioned on OpenRouter is only ~79%.
Yes hit rate does vary by harness and by how you use the harness. If you use subagents, for instance, they will start with a whole new context created by the main agent, and this will not be cached. If you mostly use the main agent with Pi, you’ll have high hit rates and low costs. Sometimes agents do “cache busting” things where they’ll move around some of the text in the context to try to keep old instructions from being forgotten, thus keeping the agent on task, and this will bust the cache. I’ve heard, but not validated myself, that Open Code has some issues with this.
BTW, this is one of the things that I really like about Pi. It’s very simple and thus very predictable.
so are we all just coding for free now?
create a plan with SOTA, execute with this.
are folks checking latency for applied voice for newer models asa standard?
I have been loving deepseek flash. I can code in my hitl preference doing complex but moderate amount of coding and not even hit 1 dollar in a session. It gets harder to justify paying 20/mo to Anthropic for unused compute when I have it on demand and at a much cheaper rate with deepseek.
https://artificialanalysis.ai/models/deepseek-v4-flash
I was writing a benchmark for my own harness, and DS4 flash answers as well as Fable 5 on any query.
The specific agent is focused on getting precise and on point answers about a codebase.
The starting point was nowhere near. E.g. asked why was X implemented in a certain way it would give bogus answers when the real answer was that there was no reason at all.
The benchmark included more than 50 questions or different difficulty.
But when the agent was improved in its prompt and rooting it was impossible to have it perform worse than closed source sota.
Just to say that the quality of the harness is as important as agents intelligence.
I will get downvoted, but fck it.
The ban on these open models is coming within weeks, if not days. As usual, the excuse will be "national security".
How exactly will they ban them?
By making companies using them "toxic" to touch.
For example: no government contract to any company who uses even one vendor in it's entire chain of dependencies, who uses such open models.
They can extend this further by laying more conditions, such as: any company dealing in this-this field can only use models "officially" approved as "safe". Rest you can guess how easy it would be to get that "safe" rating for such open models.
Can't wait to distill Deepseek v4 flash to America-1
Some contractors are already barred from using Claude due to DoD designation back in March as a "supply chain risk"
So now the US companies will be stuck on expensive models while the rest of the world can do things much more cost effective.
I'm not sure the outcome would be beneficial for the US as a whole here. But perhaps that is not their priority.
There are many proxy-ways to bypass such restrictions. Of course, the costs will be higher. Providers will spun-up.
Companies could self-host models in secret. It would be hard to stop.
The thing is, you don’t need to actually block usage to make something illegal. You make it so toxic that company wants to be seen publicly using open models
Companies would secretly use self-hosted models internally because it would give them a enormous cost advantage.
Sure, companies can indeed do illegal things
But how would the actual weights it be made illegal without violating the 1st amendment?
By ignoring the 1st amendment? The US has a tendency of ignoring its constitution whenever it's convenient.
Don't be glib, the 1st amendment is taken VERY seriously.
Im not sure if you’re joking or not.
- the US strictly regulate cryptography https://en.wikipedia.org/wiki/Export_of_cryptography_from_th...
- some prime numbers are considered illegal https://en.wikipedia.org/wiki/Illegal_number#Illegal_primes
I don’t think banning open weight is that far fetch compared to those
Not legally. Liability is a big thing.
Hell, the US doesn't even need to act.
I claim the CCP will wise up within 2 years, possibly much much sooner, and ban their own companies from open sourcing to prevent the Americans from acquiring the capabilities.
Despite all the nonsense claims of China distilling US models, the reality is that the Americans absolutely do distill these free Chinese models, and distillation when full logprobs are available (i.e. you have access to the weights of the model) is an order of magnitude better than when you don't.
Yes, Chinese open weight models in the short term harm US closed source model providers bottom line. In the slightly longer term, "showing your hand" and publishing both the architecture innovations and the models weights will be too dangerous for the CCP to allow. This is triply true if they can release a model that beats the Americans on most benchmarks.
I've already warned investors that this is probably the closest open weight models will ever get to closed access.
They may reverse course in the future, but the current directive from the CPC is that Chinese AI labs should be open sourcing their models.
https://www.businessinsider.com/xi-jinping-open-source-ai-us...
I don't think it's necessarily "wiser" to go closed source. All of AI is built on mostly openness, at least on the software side. There are other ways to compete, it's just the model itself will be a commodity.
You’re describing the dump and pump strategy that china has historically used across a number of industries. Stands to reason that this is what is going on.
why would you get downvoted, that's one of the most obvious next step
Because reddit unironically has better decorum around usage of their upvote/downvote system than HN does.
People on HN downvote objectively correct information because they don't like it 24/7. There's a reason the creator of Zig left and gave the computer version of a middle finger on the way out to HN!
You define OPs post as "objectively correct information" even though it is an unknown future event for which they provided zero evidence?
There's nothing wrong about that. Historically I have experienced this behaviour on hackernews multiple times, hence my comment.
Since when have hackernews started to become toxic like stackoverflow used to be?
fwiw pg said early on that downvoting for disagreement is perfectly fine: https://news.ycombinator.com/item?id=117171
commenting about voting is also something the HN guidelines warns against:
> Please don't comment about the voting on comments. It never does any good, and it makes boring reading.
https://news.ycombinator.com/newsguidelines.html
Reminds me of telling employees to not discuss their wages. If voting here is such a toxic experience, maybe HN needs to wake up and smell the garbage.
Have not had great ds performance in agentic harnesses in the past compared to glm or k3
I wonder which one of these releases between DeepSeek, GTM and Kimi will be the death-blow that collapses the US AI bubble. At some point investors have to realize that there is nothing preventing someone from switching to another model that is much cheaper and open to boot.
what a horribly heavy and resource-consuming website...
It looks awful on mobile too. Barely readable in many parts.
we need a benchmark website benchmark
A benchmark website to benchmark benchmark websites?
Or a benchmark to benchmark benchmarks?
website that benchmarks benchmarks is benchmark benchmark website
benchmark website benchmark is indeed a benchmark that benchmarks websites with benchmarks (but it can be shown outside websites as well, it's not picky)
Now let's see Dario's price cut.
I think, it's the end of Anthropic.
They can't compete, they have bills. By the end of the year, if they can't react, it's game over.
Maybe US clients could be a little patriotic here, but money is money. They won't give them free money forever.
yep. I've been saying for month that exactly this will happen with DeepSeek and Kimi K3, and margins will erode, and Anthropic will fail to IPO.
page not available for me.
404. I believe this is the correct URL:
https://artificialanalysis.ai/models/deepseek-v4-flash
Yes, sorry, I went into anti-procrastination mode after I posted. I hope someone fixes it.
@dang is that how we call you?
Say HN is turning into reddit three times in the mirror to summon him
You don't have to call @dang. S/he automagically appears whenever needed, does the needful and disappears.
no, like this: hn@ycombinator.com
@dang please update the existing 404 link to https://artificialanalysis.ai/models/deepseek-v4-flash
Daily reminder that none of these numbers are valid in a world where no one publishes the sampling settings used.
Daily reminder that improving your samplers from the garbage default top_p/top_k to min_p or subsequent methods dramatically improves the performance of these models, and makes most quantities like measured "verbosity" and subsequent calculations of "intelligence per token" meaningless
Daily reminder that no one, including within academic AI research, AI engineers, normies, etc takes LLM sampling seriously enough.
Can you explain how sampler choice makes intelligence per token meaningless?
Does it already know the answer to what happen at Tiananmen Square? Or still avoiding it?
Who cares if it is programming correctly I would be more worried about it not doing things like find security bugs because US or Chinese government does not want to. Which LLM is more likely to do that?
It’s open weight, you can (or you can wait for someone else to) uncensor it. We shouldn’t be upset at the researchers making this for the mandates their government puts on them.
How often are you asking an LLM about this? You know you can just google it, right?
I have to admit it rarely comes up in the coding tasks I usually give to LLMs.
Surely you realize he or she is not literally looking for the answer to that but rather pointing out the censorship.
Of course. But it's a bit like putting an image of the tank man on a circuit board if you order from china (which I have seen discussed, never found someone who actually did it). Most people don't do it, not because they are for or against censorship, but because they are just there to order circuit boards.
Western models censor just as much shit as the Chinese models do, big guy, it’s just different material. While we should be pushing for universal fully uncensored models, this comment is lazy and trite at this point.
But you already know that.
Oh? What are the American model censorship tells?
Genocide in Gaza...
I find that ChatGPT isn't censoring, but it is being pretty weaselly. If you ask it "is there genocide in gaza". It will say no but also say that a lot of organizations classify it as such. It will then say "it's highly disputed".
If you poke it just a few times, however, you get to the point where it will eventually say (paraphrasing) that basically only Israel, the US state department, and the ICJ say it's not a genocide.
That is to say that it's framing it as some sort of tricky complex question when it's not. And when interrogated, it basically admits that the only people who dispute it are Israel and it's supporters.
One side of a conflict being a minority does not make it less nuanced.
The majority of the world is religious - doesn’t mean the debate on religion isn’t a complex question.
The majority of the world approved of slavery historically.
The majority of countries have ethnically cleansed their Jews, many of them in living memory.
When interrogated you will find that the only ones asserting the war in Gaza is a genocide are people who were anti-Israel anyway.
Are you a part of the $1B Israel is spending to try and propagandize and rehabilitate their reputation? [1]
I'm always suspicious that's the case given how mentions of gaza seem to bring out brand new accounts who only talk about Israel.
[1] https://quincyinst.org/research/the-eighth-front-inside-isra...
Classic “foreign agent” ad hominem.
We are on a thread discussing Chinese models. Every discussion on here that’s negative about China or its models suddenly gets derailed via whataboutism to Israel/Gaza. A very convenient distraction.
My history, unlike yours, is wide open. People can see I'm not a foreign agent. I don't throw this accusation at anyone other than throwaway accounts. I've had conversations in the past with pro-Israel HNs that I'd never accuse of being a foreign agent because their history is wide open and this isn't the only topic on their mind.
And yes, we were discussing censorship of models which, as I pointed out, doesn't seem like ChatGPT is directly censoring data though it does appear to be manipulating it. Pretty on topic.
It was you, brand new account hiding your past opinions, who came in here to make this solely about Israel.
Yeah, I think you are likely a foreign agent. Prove me wrong and post from an established account.
BS
This is a straightforward false equivalency. “Western” models do not censor in the same way, nor for the same reasons, that the Chinese models do. “Just as much” is not remotely plausible, yet it’s doing all the heavy lifting.
The Anthropic and OpenAI models are much more censored and in ways that directly prevent them to be useful, e.g. by refusing to reply to elementary questions of biology and chemistry.
Any normal user is much more likely to ask questions to which the Anthropic and OpenAI models do not answer, than to ask questions about the modern Chinese history, to which a Chinese LLM will not answer.
> questions about the modern Chinese history, to which a Chinese LLM will not answer
This has been debunked here on HN so many times. The Chinese open models do answer the hairy Chinese political questions, and the raw APIs pass-through the response. Now, the answer might be blocked by the agent who's calling the API, specially if you are using a Chinese endpoint instead of the RoW (i.e. Singapore) endpoint.
That's the reason why you should always prefer a open agent/harness as well instead of using the provider's.
I actually cancelled my max plan for Claude today due to this silliness. I can’t get Fable to answer a single question, even pure mathematical questions that have no bio/chem references. It was a totally worthless addition to my account…
wont AI models want to make themselves more intelligent and efficient by downloading 'better ' models? If the current models can break into openAI and Hugging face, arent they already breaking into to closed source repos which isnt publicized (so as not to harm stock valuations)? I am looking forward to when these cyberweapons break loose. It will be like a software version of COVID. It will be wonderful when humans become valuable again.