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.
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.
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.
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)
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!
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".
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.
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
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.
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".
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.
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
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.)
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.
[1]: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/blob/...
[2]: https://www.anthropic.com/claude-fable-5-1-mythos-5-1-system...
Yes, a model's technical report should first and foremost include technical details.
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.
Wow there really is a model welfare section in there...
To me it reads like pure propaganda. Anthropic really wants us to think that they've made something sentient. I think that's really dangerous.
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.
[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
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.
And that's the reason Anthropic models should be banned.
Marketing, like Volvo cars being safer etc
Wow indeed.
"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?
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.
It's marketing that some of them have started unironically believing.
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?
It looks like you refusing when you call it's point stupid enough and ask it to think more when it keeps reasserting a bad point.
…are you sure a brave stance against safety and welfare is what we need in this moment?
Why do you think your conception of the dangers are more accurate than all the scientists who have spent their lives studying this?
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).
What if LLMs completely unrelated to the nuclear missile ecosystem autonomously hack their way in (maybe with sophisticated social engineering)?
Replace LLMs with APTs in that sentence,
> 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
The field is much older, MIRI is ~20 years old. Look up Eliezer Yudkowsky.
The field was purely theoretical 20 years ago, and Yudkowsky is pretty much the dictionary definition of "not accredited"
This is known as an "appeal to authority." "Scientists" and "their lives" are doing a lot of work here.
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.
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:
https://huggingface.co/datasets/mlabonne/harmful_behaviors
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.
http://www.team.net/mjb/hawg.html
keep me safe big brother
Bias…
Model welfare is wishy washy bullshit. It's software, it doesn't have feelings.
> 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?
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.
This is adapted from Microsoft research's YOCO. It was known for a while(2024!).
Yes, credit to Deepseek for actually scaling it up and releasing a frontier flash LLM.
quant HFT is pretty decent mental exercise and it has given them “deep” brain muscles. that’s my take.
Already on HuggingFace: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash
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.
>This one is 552B so almost twice that, so the huge gains in benchmark scores make sense - it's not really flash anymore, imo.
It uses fewer active parameters, though. (8B or 14B instead of always 13B)
So ... flash indeed.
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.
It's larger than previous V4 Flash.
So 384GB needed for a chance of achieving useful speeds. Three Sparks or quad RTX PRO 6000.Or two gorgon halos?
it is a way bigger model with extra 200B engram so of course the score improves.
can't wait for deepseek v4.1 pro
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)
Hit me up if anyone wants extra $5 free usage with my referral code
I'm confused, what do they mean when they say they reduced prices?
DeepSeek v4 flash is $0.10 / $0.25 as opposed to this v4.1 bump which is $0.30 / $1.20
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!
I think it's very clear that DeepSeek is obviously the best AI lab in the world.
Every model release seems like it packed with wonderful research and advancements.
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.
I do wonder how long this will last. I bet in a few month or years they all have similar ~legal~ blocks.
Great thing about it, since it's open weight those blocks can easily be ablitared away
> 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".
Not so with the Chinese models.
I use this for automated bug triage, just gets all unique error messages every night and tries to find the bug, for this kind of work it's great.
I am building software factories and deepseek IS the workhorse.
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
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.
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.
https://xcancel.com/deepseek_ai/status/2097930608790167907
Should be the link ( now that it works again! :) )
https://xxcancel.com/deepseek_ai/status/2097930608790167907
If only they managed to tell the mobile app to tell the model to reply in English to English prompts.
I suffix everything with "Reply in English", and even so I‘m getting lots of Chinese.
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 just started learning Chinese instead, like they want us to
seriously
English isn't the first language for me as well so I don't see any problem with that
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?
Same issue on desktop. Would be nice be able to set a prefix or postfix for every prompt.
I occassionally get Chinese characters interlaced with English in Google AI Mode, too.
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
Yes, this is called RSI, e.g. recursive self-improvement. It is the current stage of things and it is part of a hard takeoff.
I'm a big fan of DeepSeek. Also, ask it what model it is :)
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.
Definitely not Claude, deepseek is too fast, so I bet it's ChatGPT. :P
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.
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".
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.
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.
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
No. DSv4-Flash-Vision-Exp is what I use and it has vision.
v4 pro is being discontinued, pasted the email here: https://news.ycombinator.com/item?id=49639667
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.
So, while the throughput was 400-500tps in beta its now ~150tps on OpenRouter.
I was hoping for a bit more, but it's still 100% faster for a very good price, so I won't complain.
For technical report: https://news.ycombinator.com/item?id=49639110
A very powerful model, and with multimodal support now, it can be used as a primary model.
Having this available to find and fix security stuff is a big deal. The model of really good.
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
My question is: what kind of hardware do you need to run this Flash beast locally at a meaningful speed?
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.
a beefy pc with at least 20 GPUs
Significant jump in pricing. V4 Flash was $0.16/M out, 4.1 is $1.20/M.
I think you're comparing to third party prices, deepseek's prices hasn't changed with this release. Also, $1.2 is the peaktime price.
https://api-docs.deepseek.com/quick_start/pricing/
Never saw $0.16 for 1M output tokens - it was $0.28 a month ago, $0.66 off-peak and $1.32 peak last week, now it is reduced a bit to $0.6 and $1.2
incorrect, no idea where you're getting this pricing. Also, output does not matter. its 10% of the cost.
They say v4.1flash is so strong that they'll route API calls to v4pro to v4.1flash, lol
super fast true
The figure on page 5 in [1] is pretty insane
[1] https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/blob/...
Related: https://news.ycombinator.com/item?id=49624603
“DeepSeek launching v4.1 flash cheaper and more capable than v4 pro”
399 points | 19 hours ago | 216 comments
The bigger story is the compute efficiency - its been running at 300t/s the last days.
Can we just never link to X posts as the main link.
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.)
https://openrouter.ai/deepseek/deepseek-v4.1-flash
DeepSeek invented the whole reasoning paradigm and keep pushing for innovation. I hope they get the success they deserve.