DeepSeek-R1-Lite Preview Version Officially Released
Posted by nekofneko@reddit | LocalLLaMA | View on Reddit | 107 comments
DeepSeek has newly developed the R1 series inference models, trained using reinforcement learning. The inference process includes extensive reflection and verification, with chain of thought reasoning that can reach tens of thousands of words.
This series of models has achieved reasoning performance comparable to o1-preview in mathematics, coding, and various complex logical reasoning tasks, while showing users the complete thinking process that o1 hasn't made public.
👉 Address: chat.deepseek.com
👉 Enable "Deep Think" to try it now
Batman4815@reddit
That was... sooner than i thought considering OpenAI have been working on it for more than a year.
But damn these chinese labs are insane.
tucnak@reddit
It almost makes you think...
h666777@reddit
Lmao I'd be surprised if we don't get a report/rumor on one of the OpenAI/Anthropic employees being a spy for china by the end of the year. Manhattan style.
tucnak@reddit
That's a given, but I would say most important is to recognise that the Chinese have not, in fact, made progress that they like to say they did. It's paper mills all over. People should be reading the papers more, and lose their shit over every unproven Chinese "result" that gets reposted here. What's more pathetic: overfitting on public evals to capture attenion, or actually having your attention captured by shit like this? I don't know!
Just the other day, so-called llava-o1 was discussed. If you had actually read the paper, you would know that the o1 connection is made through Evaluation of openai o1: Opportunities and challenges of AGI—yet another paper mill product with 50 or so authors. They created that 280-page monstrosity less than two weeks after the o1 release. We don't know what O1 is doing, but it seems the Chinese have figured it out in the matter of days... They say their model performs well on visual benchmarks, but it's probably owing to the fact that they're overfitting these benchmarks in the first place.
rusty_fans@reddit
What ? No progress? Are we watching the same model releases ? They have like 3 labs pushing out very competitive open models, way more if you count closed ones. And many more that were at least open SOTA for a time. Qwen, Deepseek, Yi releases have all been very competitive at time of release. And no it's not just over fitting, these models are pretty damn good, they usually significantly improved on the latest llama release at that point in time.
Wow llava-o1 is shit, that means all chinese companies are shit? Not like there are countless examples of western startup's pulling this kind of shit.
Also keep in mind that they can't get their hands on the latest & greatest GPU tech due to sanctions and they're still giving the western companies a run for their money.
tucnak@reddit
I never said they made no progress. I'm sure the Qwen's of this world are at least as good as llama's, if not marginally better. That said, whether these models are competitive with Gemini, Claude, or even 4o for that matter—is straight up laughable. The only metric by which the Chinese models are "very competitive" is public evals. Their "performance" evaporates mysteriously in the private evals, and even though it's also true for 4o/o1 to a lesser extent, it's not true for Gemini, & Claude.
Even Gemma-9/27 are much easier aligned than any of the Qwen's that I tried, although the benchmarks would lead you to believe that Qwen's are like 1.5 stddev above gemma in all measures. And once again it's not a surprise to anybody familiar with the actual literature: had you actually read the Chinese papers, you would know the sheer extent of paper milling they're involved in, and you would also notice how they obssess about benchmarks, and techniques are "disposable pleasures"—the background for their ultimate goal to be perceived as strong.
smartwood9987@reddit
based if true
Healthy-Nebula-3603@reddit
China is working with AI much earlier in serious development than the USA .
LengthinessJumpy8409@reddit
even if their yes, I would say its best for the world or at least the open source community
djm07231@reddit
I believe that DeepSeek are former quant people who pivoted to AI after the Party started to crack down on the finance sector.
So it seems like it is a talent concentration difference the talent in the West is probably more diffused as a lot of really talented people work at Citadel or Jane Street instead of single-mindedly focusing on ML.
In China, the Party dictates several desirable strategic sectors which concentrates talent.
Billy462@reddit
China seems to have a lot of collaboration (and more open source) between top companies and universities. Over here there is obviously Meta being pretty open with models and research, but generally it’s completely closed off. At this point I think the secrecy is hurting western competitiveness.
Standard-Anybody@reddit
Seems pretty good at answering complex questions.
Got the "Man turns three switches off/on in a room to find out which switch controls which light bulb in another." Figured out it was by the heat of the lamp.
Figured out the banana on an overturned plate problem. The banana fell off the plate. Very good.
Failed at the coin on a thrown plate problem. Still assumed the energy to throw the plate automatically somehow transmitted to the coin. But did almost get it in it's thinking, just considered the possibility and for some reason didn't thoroughly pursue that line of thought.
For some reason it's brain damaged in talking in a conversation, so you only get the first question and then it just re-answer's it over and over again. No actual interaction possible.
Expensive-Paint-9490@reddit
Lite should be 15B parameters if it's like the last DeepSeek Lite. Those benchmark would be insane at that size.
_yustaguy_@reddit
Probably not the same size. My bet is that it's closer to the full size Deepseek-2
fanminghang@reddit
I tried R1-Lite on their website, and it’s much faster than DeepSeek V2.5. Based on the generation speed, R1-Lite is probably much smaller.
StevenSamAI@reddit
They said relatively small, so hard to guess, but I think their biggest model was coder V2 @ 236B parameters, so relatively small might be \~70B relative to this, but that's still pretty acessible.
However, the 236B model had a Lite version of that coder V2 at 16B parameters. I can't imagine it being that small for the benchmarks, so here's hoping for a 30-60B model? If it can be deployed on a 48GB card with plenty of context, that's geting affordable to run.
Flashy_Management962@reddit
just imagine if it is actually 16b, this would be the new secret open source
Rakhsan@reddit
Does anyone know how big it might be?
eggs-benedryl@reddit
Sorry to be that guy, but can anyone TLDR this? I'm unsure why this is such big news (not implying it isn't heh)
How large are these models expected to be?
kristaller486@reddit
Probably, this is the first public (and open-source in the future) replication of the OpenAI's o1 model. It's not just CoT, it's a more complex and challenging solution. Probably it's a small model (looks like Deepseek-V2 Lite, i.e., 16B MoE) that beets o1-preview on some math benchmarks. Because DeepSeek promises to release a full model weights and a technical report, it sounds great for open-source AI.
tucnak@reddit
You're right to question if this is worthwhile; there's conditioning at hand. Pavlovian response is such that "o1", or "reinforcement learning", or "Chinese" means upvotes. They don't understand what "RL" really means, so it's basically magic pixie dust to them. If you ask any of these people what RL is about, they would say "chain-of-thought something something" and that's it.
Healthy-Nebula-3603@reddit
not big ... I assume full version will be smaller than 100b and like maybe 20b
_yustaguy_@reddit
Mr. Altman, the whale has been awakened again...
ajunior7@reddit
The moat shrinks further
eposnix@reddit
It's impressive, but it really does feel like the entire industry is a year or more behind OpenAI. What are they cooking up while everyone else is playing catch up?
mehyay76@reddit
o1-preview did not come out a year ago. We're definitely plateauing in terms of actual "intelligence" performance.
This is why OpenAI is adding more bells and whistles like canvas etc instead of releasing a better model. o1 itself is very close to GPT-4 prompted to reason first
RMCPhoto@reddit
What makes you think we're at a plateau?
fairydreaming@reddit
This is not true.
ZebraLogic benchmark:
farel-bench benchmark:
I wouldn't call these values "very close". It's definitely a real progress and large improvement in reasoning performance.
mrjackspade@reddit
Yes, but what does actual evidence matter when you get all your information from Reddit comments and doom-mongering YouTube videos?
eposnix@reddit
o1-preview is the first iteration of the o1 model, first called q*, created a year ago.
Eralyon@reddit
Like O1, it is not for all tasks...
But it is definitively useful.
buff_samurai@reddit
Super impressive.
Us to China: you are not getting gpus China to us: you’re not making $ on gpus
grey-seagull@reddit
According to Semi Analysis deep seek has 50k hopper gpus.
https://x.com/dylan522p/status/1859302712803807696?s=46
Enough-Meringue4745@reddit
??? are you okay? Terrible take.
DaveNarrainen@reddit
How? Less customers == less profit?
Makes sense to me.
pseudonerv@reddit
nvidia and amd are not us based anyway.
intel: I give up, but please still give me money
Enough-Meringue4745@reddit
where do you guys get your facts? lol
cunningjames@reddit
Nvidia is based in California.
smarty_snopes@reddit
sad but true
Inspireyd@reddit
???
Whotea@reddit
US has an embargo of gpus on china
nekofneko@reddit (OP)
Official announcement:
DeepSeek-R1-Lite is currently still in the iterative development stage. It currently only supports web usage and does not support API calls. The base model used by DeepSeek-R1-Lite is also a relatively small model, unable to fully unleash the potential of long reasoning chains.
At present, we are continuously iterating on the inference series models. In the future, the official DeepSeek-R1 model will be fully open-sourced. We will publicly release the technical report and deploy API services.
BetEvening@reddit
>At present, we are continuously iterating on the inference series models. In the future, the official DeepSeek-R1 model will be fully open-sourced. We will publicly release the technical report and deploy API services.
Where did this official announcement come from? If so I'm so winning that manifold market bet
nekofneko@reddit (OP)
Deepseek's WeChat account
Dyoakom@reddit
I tried it. It's not as impressive in some of my tests as the hype would lead one to believe. It is however a massive step forward. If China had the GPUs that the West has, then I believe in a short time they are gonna get ahead in the race. They are doing excellent work.
GoogleOpenLetter@reddit
You only have to steal weights once, and then we're all even. It would be a completely viable strategy to just sit back and wait and let the West throw all their money in.
I think the reason Chinese models are so good is that they can access all of the training data without many copyright worries. It's particularly obvious when you look at the Ai video generators that seemingly popped up out of nowhere and were instantly the best in class, clearly they were trained on all the movies in Hollywood. The West steals this stuff too - but then they settle for cash afterwards. It's that Silicon Valley mantra of being better to ask forgiveness than permission.
Healthy-Nebula-3603@reddit
You know that model is still in training?
moarmagic@reddit
"it's still in training/still beta" isn't really a reason to pull punches when reviewing a product. One can only review what you have access to- sure it could get improved, but it could equally be abandoned, or made worse. If they aren't ready for it to be critiqued, it shouldn't be released.
No_Step3864@reddit
We need a strong reasoning model locally. That is the only thing I believe will truly start democratic movement of intelligence.
Dry-Two-2619@reddit
hugganao@reddit
That "this implies the speaker is female" is 100% bc of Asian language. Like Spanish we have gendered words for each nouns depending on if you're male or female.
Capitaclism@reddit
Open source?
Capitaclism@reddit
Open source?
BetEvening@reddit
DeepSeek better release their model to hugging face, I need to win my manifold market bet
https://manifold.markets/JohnL/by-the-end-of-q1-2025-will-an-open?play=true
SuperChewbacca@reddit
Llama 4 should sneak in before Q1 as well.
nullmove@reddit
I think in terms of tech, Meta can already beat o1 today if they want (same as Google or Anthropic). But whether a model like o1 fits in their lineup is the question. Even OpenAI said that o1 is an aside, and that the actual target is a fusion of 4o and o1 essence.
Meta will probably want to focus on full multi-modal first. Anthropic is probably just sitting on Opus because they want to see the looks of GPT-5 or whatever. I have zero doubt that Deepmind has AlphaProof like stuff that can blow o1, but as usual they have no product vision to bring it to the mortals.
I had a feeling that a one off STEM model would excite Chinese labs much more than say Mistral or Meta.
Healthy-Nebula-3603@reddit
Lol Appears it will be true
BetEvening@reddit
You just gotta believe
Dull-Divide-5014@reddit
This is quite bad model, hallucinates, not full deep answers.
phenotype001@reddit
This could solve the following task in 3 messages: "Using the finite difference method, derive the 2D update equations for simulating an incompressible fluid flow from the generic Navier-Stokes equation. Write a cavity flow simulation example using numpy and draw the vector field with matplotlib's quiver plot. "
I'm impressed.
olaf4343@reddit
The way he thinks reads like a severely sleep-deprived, highly caffeinated college freshman. Took 24 seconds and 6.8k characters to correctly answer the "plate on a banana" question. Haven't gotten a trip-up yet.
If this gets open sourced, I'll definitely be using it locally for internet research (if it's the 16b MoE, hopefully).
Infinite-Swimming-12@reddit
Doesn't seem to get the marble in the upside down cup question which i'm honestly surprised isn't in its training data
StevenSamAI@reddit
I did some of my best work as a severely sleep-deprived, highly caffeinated college freshman.
Ordinary_Mud7430@reddit
I tried it! I asked him several technical questions... I expressed surprise to him... And he answered something that I didn't ask him... I thanked him and he answered something that once again, I also didn't ask (at any time)... But that's fine to start 🙂
AnomalyNexus@reddit
Sounds promising. Fingers crossed pricing is as aggressive as their other models
StevenSamAI@reddit
It needs to be so they can gather enough user data to keep their models competitive.
AnomalyNexus@reddit
I doubt the average query is of any real interest for training data
hapliniste@reddit
Not the average one, but long chain of messages followed by a thumb down might be very helpful.
Every oai model start by shitting the bed after 5-10 messages and then in iterative updates they solve this. I think this is the data they need to do that.
O1-preview has this problem right now and I hope the user data they gather will be used to finetune o1, but we might have to wait some more months after o1 since using preview generations would bring the performance down.
StevenSamAI@reddit
I'd assume they rank and select.
While they probably use the model to generate specific synthetic training data, it helps to keep the training data diverse and relevant, so een simple, but high quality conversations will probably mix into the syntehtic chain of thought data.
ortegaalfredo@reddit
I love that China is saying, if you cripple our GPUs, we'll cripple your AI startups. It's a fight between titans where users win.
AnomalyNexus@reddit
Stoked for API. They consistently deliver good bang per buck
Enough-Meringue4745@reddit
Where's the weights?!
tucnak@reddit
Think; there's a reason why not a single lab in the West had released o1 of their own. It's because they're not convinced that RL approach like this is worthwhile. Everybody experiments with RL, it's just that OpenAI are the only ones to whom it made financial sense to release a "RL wonder-model."
Just the other day, so-called llava-o1 was discussed. If you had actually read the paper, you would know that the o1 connection is made through Evaluation of openai o1: Opportunities and challenges of AGI—yet another paper mill product with 50 or so authors. They created that 280-page monstrosity less than two weeks after the o1 release. We don't know what o1 is doing, but it seems the Chinese have figured it out in the matter of days... They say their model performs well on visual benchmarks, but it's probably owing to the fact that they're overfitting these benchmarks in the first place.
braindead_in@reddit
The reasoning thoughts are very interesting. Starts with 'Alright' It thinks with 'hmm', knows when it's confused and needs to backtrack, figures out it's going around in circles. It obviously 'understands'.
wojtess@reddit
Could be entropy based sampling? (entropix)
Healthy-Nebula-3603@reddit
When GGUF :)
Small-Fall-6500@reddit
DeepSeek was probably only able to partially dequant Bartowski's quants of their model, so that's why it's only a preview version for now. Once they get the right dequanting process down, they'll probably upload the fp16 weights.
/s
If only Bartowski quanted that fast...
djm07231@reddit
Makes me almost wish that the new Administration would lift the GPU sanctions.
The Chinese labs seem to be the only ones these days that open source really good models to the rest of us.
Imagine the things they will do without a crippling compute bottleneck.
Affectionate-Cap-600@reddit
Ok I'm really impressed.
Deus-Mesus@reddit
I just tried it on a "hard coding" problem.
It overthinks simple tasks, so expect a lot of error in simple operations, but when it reaches the point where that thinking is needed, it is quite good. So you can use it if you know what you are doing
Redoer_7@reddit
WTFFFFFFFF? In the future, the official DeepSeek-R1 model will be fully open-sourced. We will publicly release the technical report and deploy API services.
vTuanpham@reddit
Just test it with the prompt from o1 blog post:
oyfjdnisdr rtqwainr acxz mynzbhhx -> Think step by step
Use the example above to decode:
oyekaijzdf aaptcg suaokybhai ouow aqht mynznvaatzacdfoulxxz
It didn't get nowhere near the answer and give up
Whotea@reddit
Can o1 preview solve this or only the full o1?
vTuanpham@reddit
vTuanpham@reddit
Gave it hint that the first word is THERE and it still gave up. It just like me fr... 😔
PC_Screen@reddit
I gave it the hint that the number of letters on the decoded message is half of the letters on the codified message and it got it
ab2377@reddit
❤️
teachersecret@reddit
This is actually a pretty impressive demo based on my first tests. I'm excited to see this coming down the pipe - I wonder how big this model is? Looking forward to a public release :).
Healthy-Nebula-3603@reddit
Lol
It appears open source models of o1 level performance will be soon reality ...much faster than I expected....
I thought similar performance in the open source will be available in the second half of 2025 .... amazing
PC_Screen@reddit
Finally an o1 replication that doesn't try to get around doing the most important step which is reinforcement learning, I tried the cypher example prompt in the openai blogpost to compare how they reason and the reasoning chain from r1 was shockingly similar to o1 (r1 got it wrong but after a small hint it got it which is impressive), this is it, the way it backtracks is something you can only get with RL
Lumpy_Repeat_8272@reddit
i just used it. it is rly impressive, though the time it consumed is longer than the o1 preview. but it also provides full thinking steps that can enable many other models to improve! rly fascinating
fairydreaming@reddit
This is great, I can't wait to benchmark this.
Few_Painter_5588@reddit
Open model soon. I wonder how good the creative writing will be. In theory, having the mode being able to think should prevent the output from having lapses in logic.
OfficialHashPanda@reddit
Probably not that good. o1 also wasn't really an improvement in creative writing.
Status_Contest39@reddit
This is the killer, and very deepseek style :D
de4dee@reddit
mr altman, tear, down, this, wall
ihaag@reddit
I’ve been noticing at the time the web version had something different but it was only every now and then it would be like Claude and actually listen properly.
AaronFeng47@reddit
The thoughts process is fully exposed, so even if it's not open source, it would be very helpful for training open source models
Objective_Lab_3182@reddit
Did you seriously think that Sam Altman, Zuckerberg, Amodei, Pichai would beat the Chinese? How naive. Elon Musk is the only one who can beat the Chinese, he is America's hope to lead AI.
JustinPooDough@reddit
FUCK YES
NeterOster@reddit
> DeepSeek-R1-Lite is currently in the iterative development stage and only supports web usage, with no API calls available for now. The model used in DeepSeek-R1-Lite is also a smaller base model, which cannot fully unleash the potential of long reasoning chains.
> At present, we are continuously iterating on the reasoning series models. In the future, the official version of the DeepSeek-R1 model will be fully open-sourced.
Source: Translated from DeepSeek WeChat Channel
Express-Director-474@reddit
Can't wait to try it out!!
Flashy_Management962@reddit
If true absolutely insane
Valuable-Piece-7633@reddit
Wow, thanks for DeepSeek! A big gift!
kristaller486@reddit
I think will be useful share the announce tweet:
DeepSeek-R1-Lite-Preview is now live: unleashing supercharged reasoning power!
And some benchmarks:
Valuable-Piece-7633@reddit
Cool! No open source version?
kristaller486@reddit
An announce tweet says: "Open-source models and API coming soon!"
https://x.com/deepseek_ai/status/1859200141355536422