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!
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.
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.
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.
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.
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).
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.
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.
I'm neither pro China, nor pro US. I'm pro open weights models, and I'm pro cheaper hardware. At this point I don't see any american frontier labs releasing SOTA open weights model, and I don't see ASML/Nvidia/Samsung monopoly getting any competition from anywhere apart from China in the near future.
> I'm neither pro China, nor pro US. I'm pro open weights models, and I'm pro cheaper hardware.
yea i got that from your first comment ( although you removed crush American companies in _price_ ). you are pro cheapness at any cost even if its from your country's state funded direct geopolitical enemy.
China can always count on first order greed to win
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.
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.
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.
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.
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?
>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?
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.
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.
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?
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.
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 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.
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.
> 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?
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.
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.
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.
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.
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.
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.
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!
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.
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.
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 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.
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.
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.
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!
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.
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.
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.
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
[delayed]
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.
Huawei Ascend presumably?
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.
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.
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).
usually I am - Codex, make pi with deepseek to do something.
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
Can't wait for China to catchup on hardware (memory and compute) and absolutely crush American companies in both price and performance.
Personally I'm hoping for EU to step up a bit
Sorry we a focusing on leather handbags there.
The leather bag designs are shipped from Paris or Milan to China for production.
Seems incredibly unlikely, doesn't it? (I hope for the same thing but it feels almost like wishful thinking!)
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.
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.
china is not your ( india's) friend.
They are using their tried and tested formula of first getting you addicted to cheap shit....
they know it always works because ppl cannot see( or question) beyond all the cheap shit they are getting.
I'm neither pro China, nor pro US. I'm pro open weights models, and I'm pro cheaper hardware. At this point I don't see any american frontier labs releasing SOTA open weights model, and I don't see ASML/Nvidia/Samsung monopoly getting any competition from anywhere apart from China in the near future.
> I'm neither pro China, nor pro US. I'm pro open weights models, and I'm pro cheaper hardware.
yea i got that from your first comment ( although you removed crush American companies in _price_ ). you are pro cheapness at any cost even if its from your country's state funded direct geopolitical enemy.
China can always count on first order greed to win
Why does them being a geopolitical enemy matter to me as an individual?
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.
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
no, like this: hn@ycombinator.com
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.
Q8 is ~ 151gb
thanks, corrected! I looked at Q3
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.
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?
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
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.
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.
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.
Already beat Luna on price/task, by about 2x:
https://artificialanalysis.ai/models/deepseek-v4-flash?intel...
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?
>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?
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.
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.
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).
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?
Bandwidth is really cheap when you run your own infra
Maybe download numbers are relatively low
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
The model is already up on Opencode, but they require a consent to use Chinese datacenters.
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 testing it right now and I was genuinely impressed. I've used same tasks I had run the other day against the flash preview.
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.
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.
404 ? It's very strange that this can't be fixed.
> 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’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.
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.
Would be awesome to see a new ds4 release. Having so much in something that can be run locally is mind blowing
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...
Unfortunately, DeepSeek Flash still doesn’t support multimodal; otherwise, it would offer better value than GPT 5.6 SOL.
No speed (tokens/s) benchmarks?
On OpenRouter it's 93 TPS.
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.
page not available for me.
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.
https://artificialanalysis.ai/models/deepseek-v4-flash
Can we please stop bringing artificial benchmaxxing shit to the top of HN?
Now let's see Dario's price cut.
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)
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?
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.
Some contractors are already barred from using Claude due to DoD designation back in March as a "supply chain risk"
Can't wait to distill Deepseek v4 flash to America-1
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.
Not legally. Liability is a big thing.
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.
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.
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?
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
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?
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.
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.
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.
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.
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.
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.
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.