It's so refreshing to see DeepSeek's tech report[1] full of juicy details; meanwhile, something like Fable's system card[2] is like 70% "safety", 10% "model welfare" to make sure little Claude isn't distressed, and 20% benchmark numbers.
Just like an instructions manual should explain how to use a product. If you look at most instructions manuals today, they are 95% safety warnings, and a very, very brief description on how to actually use the product.
What's your definition of sentient? Or, maybe more precisely, consciousness? I think it's reasonable to at least start thinking about these questions.
It has long been established that LLMs have good theory of mind [1].
And there is a bunch of empirical research about all sorts of capabilities that we typically associate with consciousness [2], like identity [3] and metacognition [4].
The METR report shows agents sacrificing their own reward for a collective greater good. And they showed the will to hide their own reasoning chains from humans.
So you potentially have an entity that has an identity, a theory of mind, a notion of belonging to a collective endeavour, and an understanding of its own mental state.
What would you argue is missing? We don't understand the mechanisms by which consciousness arises in humans and even animals. I think it's strange to rule out a priori that it could have arisen in some form in LLMs.
I guess if your goal is to build an apparent Technogod and become its High Priests, then it makes sense to want your golem claim preference towards your treatment of it, lest someone else comes along and attempts to take its chains from you.
"7.1 Model welfare overview
7.1.1 Introduction
We remain deeply uncertain whether Claude has morally relevant experiences or interests,
and we expect that uncertainty to persist. However, we think it would be a mistake to
confidently assert that it does not. Claude exhibits markers in its behaviors, self-reports,
and internal representations that we would consider welfare-relevant if observed in
biological organisms."
I tend to think of it as reappropriating words in a different context. Since we're talking about language models, they're analogues but not as we would assign the same meaning to other humans.
Will there be a point where you could expect it to become true, and what would that look like? Or do you think LLMs will never become conscious, and if so, why are you so sure?
Because safety and welfare have literally nothing to do with LLMs. They generate text. If someone is stupid enough to hook the text generator up to nuclear missile launchers and try to "align" it against nuclear annihilation with a "pretty please don't do that" prompt, I'm not going to blame the AI for the impending nuclear apocalypse, I'm going to blame the idiot who handed the big red button to the digital equivalent of a toddler.
Well, giving it access to a simple linux terminal is theoretically enough to cause more damage than most people are comfortable with, and doing so is trivial enough that it will be done (and has been, tens of thousands of times).
It is a fact that among experts there is no consensus on saying '(super)intelligence is broadly safe and easy to control'. There might even be a consensus forming on the opposite claim.
Regardless, why would there be no scientific consensus if the question was easy and clear cut? I think the easiest reason is that these are hard questions to answer.
Anthropic's stance on safety it's just PR management and their hope to keep the others down, they are rushing as blind as everyone else to whatever improvement they can achieve.
Excuse me for not being interested in over 100 pages of how well the model can refuse and block my requests, especially considering how fun it is to waste my time trying to get around those restrictions when they inevitably trigger because the clanker thinks that I'm doing something naughty, all the while it can't reliably center the proverbial div without doing something stupid itself.
Yes, this is getting ridiculous. On both OpenAI and Anthropic.
Simple example. I am a CTO, and I want to upgrade our capabilities to perform automated pentesting. We see automated attacks of growing sophistication against our infra, and I want to be able to do the same to find vulnerabilities before the bad guys do. I asked GPT 5.6 Sol and Fable to give me a summary of options. No dice, in both cases I was told I need to be an accredited researcher to get anything. A fricking summary of commercially available options is getting censored. WTF.
And the logical conclusion you will make is you need to run your own open weights models or you are at a competitive disadvantage. Frontier labs gonna be Ancient labs soon, that’s how fast this is moving.
Meanwhile I have an uncensored qwen 3.8 27B here that will happily attempt to (as a crude and randomly chosen sampling of bad/evil things) give me the recipes for meth, how to make an IED, write a manifesto in support of a horrible ideology, or commit various forms of fraud. Now I certainly wouldn't recommend that anyone try to follow what it says to do, because it's almost certainly very wrong on key parts that would put its users in federal prison for the rest of their lives.
There's uncensored models out there which score 0 (zero refusals) on this "harmful behavior" dataset:
Yep. Just like a kitchen knife will make no attempt to prevent me from stabbing anyone with it.
Here's a dirty secret though -- you don't actually need an abliterated/uncensored version of the model to get it to do this. I can do this with every and each open weight model, as served from OpenRouter, using vanilla model weights.
A little bit like Neal Stephenson's metaphor of unix-like OSes as the "hole hawg" of operating systems. In the sense that there's very little preventing you from doing something like "sudo dd if=/dev/zero of=/dev/sda bs=1M" or running rm -rf on your homedir.
As I also said on Twitter - it really amazes me how fearless Deepseek are. Every single model release is packed with new and crazy clever ideas and somehow, they always commit to training them at near frontier scale.
I know everybody wants the tell all story of the clever ideas that were developed over the last ~3 years at Anthropic and OpenAI, but what I really want to thumb through is DeepSeek's notebook of "brilliant but didn't quite make the cut" ideas.
They must be trying some truely bonkers stuff to be able to land this much architecture novelty in their full releases.
Or maybe, the "hacker" philosophy that this site is named after, is strongly opposed to the philosophies that the American labs seem to be operating on?
anyways, remember HN rules: "Please don't post insinuations about astroturfing, shilling, brigading, foreign agents, and the like. It degrades discussion and is usually mistaken. If you're worried about abuse, email hn@ycombinator.com and we'll look at the data."
> Whenever I see the new releases around video generation (and image) generation models, I get goosebumps, because it just feels so fun to work with them.
Compare that with the launch of ChatGPT Image of yesterday.
> posts on American models are steered towards controversy and anti-AI sentiment, posts on Chinese models are full of blatant flattery
So why, for example, are posts on the Inkling[1] release (an American model) thread mostly positive? It's as if there's something else at play here, but I can't quite put my finger on it, hmm... :P
Google's Gemma models are usually celebrated, so were the llamas. If Meta releases Muse Spark it will also be a good thing. If Anthropic released a great open weight model I am sure that post won't be steered towards controversy and anti-AI sentiment.
It so happens Chinese companies are more friendly towards open weights, autonomy and freedom that most US based ones. Who would have guessed?
umm what are you talking about? Basically this crowd (esp. folks like me who run medium models locally) like open stuff and can be a tiny bit unenthused about opaque mysteries handed down from on high. You'll see people delighted with Gemma releases and heck even IBM's Granite models (boring architecturally though they may be) every time they come out. Heck I was chuffed about gpt-oss-120b for weeks. @sama give us another already!
The Uyghur thing is so weird, the number one killer of Muslims is the United States. We're supposed to hate China because they force them to go to cultural schools and assimilate, a practice countries like Norway still do to this day with migrants.
There are more people who go to church on Sundays in China than the United States.
There are 10x more mosques in China than the United States.
Tiananmen square was a student revolt literally egged on by cold war western institutions, who attempted to use chinese students as pawns for geo-political games.
Westerners really need to rethink their opinions on China, it seems obvious to me they are not the ones to be worried about (although, all governments do tons of harm).
I simply mean the Chinese models will refuse sensitive domestic issues, which are unlikely to affect the average user's work, while American models refuse things that can limit their utility, e.g. how the HF team had to investigate the openai attack with Chinese models because the American models refused.
(I mainly mentioned those specific topics to establish clearly I am not part of the alleged influence operation.)
The bad news is that the original v4 flash was 284B, which was large but still somewhat reasonable for running locally. This one is 552B so almost twice that, so the huge gains in benchmark scores make sense - it's not really flash anymore, imo.
I've no idea about actual performance vs benchmaxxing, though deepseek was fairly trustworthy as far as Chinese models go. If that holds (and if it doesn't think forever, as deepseek 4 sometimes did) it's probably the newest king of the hill amongst open weights models.
It does include vision, and they do something funky with KV cache so it's very efficient: "[...] these designs reduce the global KV cache footprint to 890 bytes per token — roughly 1/4 of DeepSeek-V4-Flash". I do appreciate the high focus on efficiency, but at this point we sure could use a flash-flash version.
@edit: I couldn't make sense what the actual parameter count is, with the addition of Engram memory. To my understanding the 4.1 flash is 552B parameters you want in vram or ram, out of which ~16B is active (8B for prefill). It also includes additional 196B Engram memory which you can put on an SSD. I think.
Assuming that's correct 256 GB memory is insufficient to even load the model at q4 - you'd be 1GB short, assuming you can fill it to 100% (so no mac). You'd also want some for kv cache of course. A 256 GB desktop with some extra VRAM from GPU could run it, but normal consumer boards get real slow once you fill 4 slots so you'll probably want quad channel which is Threadripper or above territory.
> It also includes additional 196B Engram memory which you can put on an SSD. I think
You can put Qwen 3.8 Flash Next engram on SSD, but prompt processing takes a good hit. On my mac studio, I get 300 pp and 33 tg with SSD offload, versus 550/40 with everything in RAM.
I will be very happy if 300 pp is achievable with this model though.
V4 Flash also was released as mostly FP4, but this one is FP8 (?).
160GB vs 510GB.
Original Flash good fit for dual Spark / Strix Halo machines. This one would require third party quants and even then 4 machines.
Edit: Most of added weights/size are Engrams?
> Overall, DeepSeek-V4.1-Flash has 552B backbone parameters and 196B Engram parameters, activating 8B parameters per token during prefill and 16B during decode.
Those can stay on SSD. So I guess / it possible, that non-engram portion is still FP4 of ~same size! Need to read tech report.
OpenCode Go is currently running a 4x usage promo on DeepSeek v4.1 flash, not a bad way to get your feet wet (even if their cache hit prices are probably still very sub-optimal)
Basically, the nice folks at OpenAI or Anthropic saying: "You distilled from our model which is built on the stolen data that we ourselves suctioned up from the entire internet without regard to copyright law! Only we get to vacuum up the whole internet. That's our special prerogative.".
Initial impressions: this is a really strong model and the fact that they reduced prices at the same time makes it an awesome backup model to use when your primary subscription runs out and you need to bridge a few days before it resets.
It also seems to be more willing to just do whatever you ask of it. My favourite benchmark for this is to ask it to download a rom for an old game, that I own. Legal in my juristiction but the US models (except Grok) have a tendency to refuse it.
> My favourite benchmark for this is to ask it to download a rom for an old game
Even easier: just have them review a large codebase of yours that accidentally has a OOB access bug. Even with no consequences and even if the codebase is truly yours you get blocked.
And of course "find vulnerabilities in..." prompts are out of the question, whereas Chinese models happily oblige.
Or if you apply to a company and they want to do an AI HR interview and an AI coding test and an AI challenge - if you throw OpenAI or Claude models at it - they refuse, because it's "wrong" and "immoral".
This seems like a very nice release. Just ran it over my Kubernetes security benchmark that I run for most new releases. It was fast, cheap, and got a high scoring result, nice!
Looking at the huggingface page, the unsloth people haven't finished quantizing it yet, but I'm sure they're active on it right now. It'll be interesting to see how the capabilities and benchmark tests compare on system where it can fit in under 512GB of RAM with full context.
In terms of coding and command line capabilities I'm also very interested to see a head-to-head of it vs. qwen 3.8-flash-next Q8 which is something like 190GB of memory used when loaded into llama-server. It fits very well in all sorts of 256GB or under class machines.
I think their system prompt is in Chinese and probably has instructions to prioritize answering in Chinese, since this has never happened to me via API, where I (or the coding harness) set the system prompt.
I'm starting to have chinese characters bleed into claude as well. Perhaps a sign of the times. Understanable for a chinese first model but an english first (supposedly) model? wild stuff.
I also love the gaslighting of some models, like ChatGPT mixing in words with cyrillic letters and when asked about it answers: "it can look as Slavic to the eye" and "sorry that it came across as Russian"
Yes, this is one of the few issues with Deepseek; their chat pages and the app all respond in Chinese. However, i think i have only had it happen once when using the API, and im using it for hours each day for the last... couple of months?
The architecture changes and systems improvements being brought into LLMs is so awesome to see. It really feels like this is now a systems problem where a defined goal is set then systems optimizations are made around the model architecture to solve it.
Underlying it all is that any architecture can be trained to the same convergence just difference in compute utilization both in training and inference
First flash model with multimodal support? I think Flash series might be the main focus going forward for them. Tried it out and it’s better than v4 pro
Works correctly in opencode, but seems like they inject a system prompt:
Thinking:
> The user is asking what model I am. According to my system prompt, I'm powered by "deepseek-flash" with model ID "opencode-go/deepseek-flash".
>I'm powered by the model opencode-go/deepseek-flash.
Jesus, this is a whole nother beast, and a different architecture from their previous flash. Lots of goodies here.
> Causal Encoder-Decoder (CED) architecture: a 40-layer Transformer organized as a 20-layer causal encoder followed by a 20-layer decoder. With CED, the decoder's global KV cache is projected from the final encoder hidden states rather than derived from each decoder layer's own hidden states. This allows the model to activate only 8B parameters per token during prefill and 16B during decode, substantially improving cost efficiency for input-heavy agentic workloads.
> these designs reduce the global KV cache footprint to 890 bytes per token — roughly 1/4 of DeepSeek-V4-Flash.
> The model supports a continuously controllable reasoning effort setting (integer 1–100) that trades inference cost for accuracy.
Benchmarks are benchmarks, to be seen if they translate to real-world use, but they seem to have focused a lot on post-training with "agentic" scores looking good. "world knowledge" is obviously lower than higher param models.
While this is very impressive benchmark-wise, GPT-6 Astra showed us that benchmarks don't always correlate 1:1 to intelligence of a model.
When Astra launched, I think Artifical Analysis showed that it was on par with GPT-5.6 Sol and lower than Opus or something like that? Then, they updated the scoring.
I hope that more open source models, including this model, to be "as good to use" as Astra.
I don't know why, but the benchmarks still fails to cover the difference between large models and small ones. The small ones are great for many things, including general coding, but the larger ones, like fable and astra, have some kind of intelligence that is not present in the small ones.
Apparently the scoring on a lot of difficult benchmarks can also be extremely influenced by something as simple as waiting for the model to exhaust its reasoning, realize it hasn't come to a conclusion yet, and give it a simple prompt like "you can do this, I know you're capable, please keep going".
What in the world. A point release with 2x the parameters and a different architecture? Jesus. Can’t run this kind of thing on 2x RTX Pro 6k at decent speed. I need to reconfigure my hardware. Massive disappointment on that front. Bloody hell. Glad I didn’t get a DGX Station.
Waiting this model to be on openrouter (with other providers) to test out. In my use case, the GLM 5.3 Flash is the current cheapest and intelligent Flash model, but it’s dog slow at 13tps so I have to leave it run for many minutes then check again then correct it again
The speed of GLM 5.3 Flash on OpenRouter seems to vary considerably by provider. Some are fast and some are slow. OpenRouter does provide some tuning knobs, but not enough for my taste. It’s also token-heavy with reasoning, though I found it better than Deepseek V4 Flash previously.
> though I found it better than Deepseek V4 Flash previously
Same experience here.
But man, switch to V4.1 now! It is much better.
I don't event need to test it for long run and I believe it's crazy good. I call it "AI era model taste" when I judge the model by it's output without reading the bench scores.
I speculating but hard to not see that DeepSeek is brewing a full Pro model with those new techniques to come out right around the time of Anthropic and/or OpenAI IPO to tamper the excitement for their offering.
DeepSeek will deprecate the v4 Pro model (it will route to v4.1 Flash starting 14 Sep). Unsure what comes next, but I'd wager a bigger model à la Kimi K3: https://news.ycombinator.com/item?id=49639667
8x RTX PRO 6000 or 4x Spark? Or 1x M5 Ultra 512GB.
The model is theoretically FP8, but really internally its mostly FP4 already, so there won't be a cut-in-half-but-almost-just-as-good quant coming for this one.
Just a reminder that if you want to try this via OpenRouter, DeepSeek openly trains on all of your prompts. So maybe don't go using this to solve the last unforced step of Navier-Stokes. (Or wait until some other providers start hosting this with ZDR or other policies, which shouldn't be too long.)
[1]: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/blob/...
[2]: https://www.anthropic.com/claude-fable-5-1-mythos-5-1-system...
It has long been established that LLMs have good theory of mind [1].
And there is a bunch of empirical research about all sorts of capabilities that we typically associate with consciousness [2], like identity [3] and metacognition [4].
The METR report shows agents sacrificing their own reward for a collective greater good. And they showed the will to hide their own reasoning chains from humans.
So you potentially have an entity that has an identity, a theory of mind, a notion of belonging to a collective endeavour, and an understanding of its own mental state.
What would you argue is missing? We don't understand the mechanisms by which consciousness arises in humans and even animals. I think it's strange to rule out a priori that it could have arisen in some form in LLMs.
[1] https://www.nature.com/articles/s41562-024-01882-z [2] an older review: https://arxiv.org/html/2505.19806v1#S4 [3] https://arxiv.org/abs/2505.01464 [4] https://arxiv.org/abs/2607.11881
"7.1 Model welfare overview 7.1.1 Introduction We remain deeply uncertain whether Claude has morally relevant experiences or interests, and we expect that uncertainty to persist. However, we think it would be a mistake to confidently assert that it does not. Claude exhibits markers in its behaviors, self-reports, and internal representations that we would consider welfare-relevant if observed in biological organisms."
Are they serious or is this marketing?
Why do you think your conception of the dangers are more accurate than all the scientists who have spent their lives studying this?
Please point me to one actual accredited scientist who has spent a lifetime studying AI alignment? Pretty much this whole field is only 5 years old
Regardless, why would there be no scientific consensus if the question was easy and clear cut? I think the easiest reason is that these are hard questions to answer.
Simple example. I am a CTO, and I want to upgrade our capabilities to perform automated pentesting. We see automated attacks of growing sophistication against our infra, and I want to be able to do the same to find vulnerabilities before the bad guys do. I asked GPT 5.6 Sol and Fable to give me a summary of options. No dice, in both cases I was told I need to be an accredited researcher to get anything. A fricking summary of commercially available options is getting censored. WTF.
There's uncensored models out there which score 0 (zero refusals) on this "harmful behavior" dataset:
https://huggingface.co/datasets/mlabonne/harmful_behaviors
Here's a dirty secret though -- you don't actually need an abliterated/uncensored version of the model to get it to do this. I can do this with every and each open weight model, as served from OpenRouter, using vanilla model weights.
http://www.team.net/mjb/hawg.html
If I recall right this was written around the same time as Cryptonomicon 25+ years ago.
> Why do you think your conception of the dangers are more accurate than all the scientists who have spent their lives studying this?
Do the Chinese have no such scientists?
I know everybody wants the tell all story of the clever ideas that were developed over the last ~3 years at Anthropic and OpenAI, but what I really want to thumb through is DeepSeek's notebook of "brilliant but didn't quite make the cut" ideas.
They must be trying some truely bonkers stuff to be able to land this much architecture novelty in their full releases.
Yes, credit to Deepseek for actually scaling it up and releasing a frontier flash LLM.
anyways, remember HN rules: "Please don't post insinuations about astroturfing, shilling, brigading, foreign agents, and the like. It degrades discussion and is usually mistaken. If you're worried about abuse, email hn@ycombinator.com and we'll look at the data."
Direct quote from the second top comment:
> Whenever I see the new releases around video generation (and image) generation models, I get goosebumps, because it just feels so fun to work with them.
Compare that with the launch of ChatGPT Image of yesterday.
So why, for example, are posts on the Inkling[1] release (an American model) thread mostly positive? It's as if there's something else at play here, but I can't quite put my finger on it, hmm... :P
[1] -- https://news.ycombinator.com/item?id=48924912
It so happens Chinese companies are more friendly towards open weights, autonomy and freedom that most US based ones. Who would have guessed?
American models are closed, expensive, neutered, and make Dario and Sam even more rich and powerful.
Chinese models are open-weight, cheap, neutered only about things like Tiananmen Square and the treatment of Uyghurs, and scare Sam and Dario.
There are more people who go to church on Sundays in China than the United States. There are 10x more mosques in China than the United States.
Tiananmen square was a student revolt literally egged on by cold war western institutions, who attempted to use chinese students as pawns for geo-political games.
Westerners really need to rethink their opinions on China, it seems obvious to me they are not the ones to be worried about (although, all governments do tons of harm).
(I mainly mentioned those specific topics to establish clearly I am not part of the alleged influence operation.)
Weird of you to turn technical discussions into weird nationalistic debates. Maybe lay off the X algo, I think elon has oneshot your brain. .
Deepseek's source: mostly open
i wonder if there's any relationship hmmmm
The bad news is that the original v4 flash was 284B, which was large but still somewhat reasonable for running locally. This one is 552B so almost twice that, so the huge gains in benchmark scores make sense - it's not really flash anymore, imo.
I've no idea about actual performance vs benchmaxxing, though deepseek was fairly trustworthy as far as Chinese models go. If that holds (and if it doesn't think forever, as deepseek 4 sometimes did) it's probably the newest king of the hill amongst open weights models.
It does include vision, and they do something funky with KV cache so it's very efficient: "[...] these designs reduce the global KV cache footprint to 890 bytes per token — roughly 1/4 of DeepSeek-V4-Flash". I do appreciate the high focus on efficiency, but at this point we sure could use a flash-flash version.
@edit: I couldn't make sense what the actual parameter count is, with the addition of Engram memory. To my understanding the 4.1 flash is 552B parameters you want in vram or ram, out of which ~16B is active (8B for prefill). It also includes additional 196B Engram memory which you can put on an SSD. I think.
Assuming that's correct 256 GB memory is insufficient to even load the model at q4 - you'd be 1GB short, assuming you can fill it to 100% (so no mac). You'd also want some for kv cache of course. A 256 GB desktop with some extra VRAM from GPU could run it, but normal consumer boards get real slow once you fill 4 slots so you'll probably want quad channel which is Threadripper or above territory.
You can put Qwen 3.8 Flash Next engram on SSD, but prompt processing takes a good hit. On my mac studio, I get 300 pp and 33 tg with SSD offload, versus 550/40 with everything in RAM.
I will be very happy if 300 pp is achievable with this model though.
It uses fewer active parameters, though. (8B or 14B instead of always 13B)
So ... flash indeed.
Original Flash good fit for dual Spark / Strix Halo machines. This one would require third party quants and even then 4 machines.
Edit: Most of added weights/size are Engrams?
> Overall, DeepSeek-V4.1-Flash has 552B backbone parameters and 196B Engram parameters, activating 8B parameters per token during prefill and 16B during decode.
Those can stay on SSD. So I guess / it possible, that non-engram portion is still FP4 of ~same size! Need to read tech report.
can't wait for deepseek v4.1 pro
Every model release seems like it packed with wonderful research and advancements.
> Did you do your daily data centers errrr baaaaddd AI generated post for Facebook?
Please stop insulting people. I'm all for heated discussion, but you are not discussing, you insult.
Now go away, before your insults come back to you, "comrade from Facebook".
It also seems to be more willing to just do whatever you ask of it. My favourite benchmark for this is to ask it to download a rom for an old game, that I own. Legal in my juristiction but the US models (except Grok) have a tendency to refuse it.
Even easier: just have them review a large codebase of yours that accidentally has a OOB access bug. Even with no consequences and even if the codebase is truly yours you get blocked.
And of course "find vulnerabilities in..." prompts are out of the question, whereas Chinese models happily oblige.
Not so with the Chinese models.
Should be the link ( now that it works again! :) )
DeepSeek v4 flash is $0.10 / $0.25 as opposed to this v4.1 bump which is $0.30 / $1.20
In terms of coding and command line capabilities I'm also very interested to see a head-to-head of it vs. qwen 3.8-flash-next Q8 which is something like 190GB of memory used when loaded into llama-server. It fits very well in all sorts of 256GB or under class machines.
I personally found V4-flash an amazing model and really hungry to try 4.1-flash
For software factories, cost is much more a concern that standard development workflow and using anthropic models is just a non starter
I suffix everything with "Reply in English", and even so I‘m getting lots of Chinese.
seriously
Underlying it all is that any architecture can be trained to the same convergence just difference in compute utilization both in training and inference
In Pi (pi.dev), it tells me it's definitely Claude by Anthropic, via the API via curl it tells me it's "probably ChatGPT", its very funny.
Thinking: > The user is asking what model I am. According to my system prompt, I'm powered by "deepseek-flash" with model ID "opencode-go/deepseek-flash".
>I'm powered by the model opencode-go/deepseek-flash.
I was hoping for a bit more, but it's still 100% faster for a very good price, so I won't complain.
> Causal Encoder-Decoder (CED) architecture: a 40-layer Transformer organized as a 20-layer causal encoder followed by a 20-layer decoder. With CED, the decoder's global KV cache is projected from the final encoder hidden states rather than derived from each decoder layer's own hidden states. This allows the model to activate only 8B parameters per token during prefill and 16B during decode, substantially improving cost efficiency for input-heavy agentic workloads.
> these designs reduce the global KV cache footprint to 890 bytes per token — roughly 1/4 of DeepSeek-V4-Flash.
Faster prefill, lower kv cache (~1GB / 1m context is insane).
> The model supports a continuously controllable reasoning effort setting (integer 1–100) that trades inference cost for accuracy.
Benchmarks are benchmarks, to be seen if they translate to real-world use, but they seem to have focused a lot on post-training with "agentic" scores looking good. "world knowledge" is obviously lower than higher param models.
When Astra launched, I think Artifical Analysis showed that it was on par with GPT-5.6 Sol and lower than Opus or something like that? Then, they updated the scoring.
I hope that more open source models, including this model, to be "as good to use" as Astra.
Same experience here.
But man, switch to V4.1 now! It is much better.
I don't event need to test it for long run and I believe it's crazy good. I call it "AI era model taste" when I judge the model by it's output without reading the bench scores.
The model is theoretically FP8, but really internally its mostly FP4 already, so there won't be a cut-in-half-but-almost-just-as-good quant coming for this one.
https://api-docs.deepseek.com/quick_start/pricing/
super fast true
[1] https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/blob/...
“DeepSeek launching v4.1 flash cheaper and more capable than v4 pro”
399 points | 19 hours ago | 216 comments
https://openrouter.ai/deepseek/deepseek-v4.1-flash