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.
Agreed on the difficulty tuning being the weak point. The interesting question is what a difficulty knob does to your own conversion rate - if it costs a real user two seconds, that is the number that eventually caps how high you can turn it.
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.
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.
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."
> 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
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 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.
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.
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".
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.
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.
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.
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?
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"
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
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
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
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...
Agreed on the difficulty tuning being the weak point. The interesting question is what a difficulty knob does to your own conversion rate - if it costs a real user two seconds, that is the number that eventually caps how high you can turn it.
Yes, a model's technical report should first and foremost include technical details.
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.
1 reply →
[dead]
…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.
2 replies →
> 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
This is known as an "appeal to authority." "Scientists" and "their lives" are doing a lot of work here.
1 reply →
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.
1 reply →
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?
Bias…
keep me safe big brother
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.
quant HFT is pretty decent mental exercise and it has given them “deep” brain muscles. that’s my take.
Oh the glazing...
Related discussion on HN - https://news.ycombinator.com/item?id=49624598
TL;DR - posts on American models are steered towards controversy and anti-AI sentiment, posts on Chinese models are full of blatant flattery.
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."
1 reply →
Not Chinese/American models. We are talking about open-weight (and their detailed tech report) and close-weight (with non-sense restrictions)
> 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
[1] -- https://news.ycombinator.com/item?id=48924912
This post doesn't even allege this...
Weird of you to turn technical discussions into weird nationalistic debates. Maybe lay off the X algo, I think elon has oneshot your brain. .
This doesn't require an influence operation.
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.
Your source: vibes
Deepseek's source: mostly open
i wonder if there's any relationship hmmmm
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!
1 reply →
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.
it is a way bigger model with extra 200B engram so of course the score improves.
can't wait for deepseek v4.1 pro
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.
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".
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.
For technical report: https://news.ycombinator.com/item?id=49639110
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.
Without a doubt, uncontested best distillers in the world.
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.
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.
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
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?
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"
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.
Having this available to find and fix security stuff is a big deal. The model of really good.
https://xxcancel.com/deepseek_ai/status/2097930608790167907
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.
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
v4 pro is being discontinued, pasted the email here: https://news.ycombinator.com/item?id=49639667
DeepSeek invented the whole reasoning paradigm and keep pushing for innovation. I hope they get the success they deserve.
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.
Significant jump in pricing. V4 Flash was $0.16/M out, 4.1 is $1.20/M.
My question is: what kind of hardware do you need to run this Flash beast locally at a meaningful speed?
a beefy pc with at least 20 GPUs
They say v4.1flash is so strong that they'll route API calls to v4pro to v4.1flash, lol
super fast true
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
A very powerful model, and with multimodal support now, it can be used as a primary model.
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
The bigger story is the compute efficiency - its been running at 300t/s the last days.
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
Can we just never link to X posts as the main link.
[dead]