3.2-Exp came out in September: this is 3.2, along with a special checkpoint (DeepSeek-V3.2-Speciale) for deep reasoning that they're claiming surpasses GPT-5 and matches Gemini 3.0
The assumption here is that 3.2 (without suffix) is an evolution of 3.2-Exp rather than being the same model, but they don't seem to be explicitly stating anywhere whether they're actually different or that they just made the same model GA.
Benchmarks are super impressive, as usual. Interesting to note in table 3 of the paper (p. 15), DS-Speciale is 1st or 2nd in accuracy in all tests, but has much higher token output (50% more, or 3.5x vs gemini 3 in the codeforces test!).
Yet tbh if the US industry had not moved ahead and created the race with FOMO it would not had been easier for Chinese strategy to work either.
The nature of the race may change as yet though, and I am unsure if the devil is in the details, as in very specific edge cases that will work only with frontier models ?
Valuation is not based on what they have done but what they might do, I agree tho it's investment made with very little insight into Chinese research. I guess it's counting on deepseek being banned and all computers in America refusing to run open software by the year 2030 /snark
It's all about the hardware and infrastructure. If you check OpenRouter, no provider offers a SOTA chinese model matching the speed of Claude, GPT or Gemini. The chinese models may benchmark close on paper, but real-world deployment is different. So you either buy your own hardware in order to run a chinese model at 150-200tps or give up an use one of the Big 3.
The US labs aren't just selling models, they're selling globally distributed, low-latency infrastructure at massive scale. That's what justifies the valuation gap.
Edit: It looks like Cerebras is offering a very fast GLM 4.6
There is a great deal of orientalism --- it is genuinely unthinkable to a lot of American tech dullards that the Chinese could be better at anything requiring what they think of as "intelligence." Aren't they Communist? Backward? Don't they eat weird stuff at wet markets?
It reminds me, in an encouraging way, of the way that German military planners regarded the Soviet Union in the lead-up to Operation Barbarossa. The Slavs are an obviously inferior race; their Bolshevism dooms them; we have the will to power; we will succeed. Even now, when you ask questions like what you ask of that era, the answers you get are genuinely not better than "yes, this should have been obvious at the time if you were not completely blinded by ethnic and especially ideological prejudice."
1. Chinese models typically focus on text. US and EU models also bear the cross of handling image, often voice and video. Supporting all those is additional training costs not spent on further reasoning, tying one hand in your back to be more generally useful.
2. The gap seems small, because so many benchmarks get saturated so fast. But towards the top, every 1% increase in benchmarks is significantly better.
On the second point, I worked on a leaderboard that both normalizes scores, and predicts unknown scores to help improve comparisons between models on various criteria: https://metabench.organisons.com/
You can notice that, while Chinese models are quite good, the gap to the top is still significant.
However, the US models are typically much more expensive for inference, and Chinese models do have a niche on the Pareto frontier on cheaper but serviceable models (even though US models also eat up the frontier there).
They're not that close (on things like LMArena) and being cheaper is pretty meaningless when we are not yet at the point where LLMs are good enough for autonomy.
I would expect one of the motivations for making these LLM model weights open is to undermine the valuation of other players in the industry. Open models like this must diminish the value prop of the frontier focused companies if other companies can compete with similar results at competitive prices.
It's awesome that stuff like this is open source, but even if you have a basement rig with 4 NVIDIA GeForce RTX 5090 graphic cards ($15-20k machine), can it even run with any reasonable context window that isn't like a crawling 10/tps?
Frontier models are far exceeding even the most hardcore consumer hobbyist requirements. This is even further
Well props to them for continuing to improve, winning on cost-effectiveness, and continuing to publicly share their improvements. Hard not to root for them as a force to prevent an AI corporate monopoly/duopoly.
I don't want to make big generalizations. But one thing I noticed with chinese models, especially Kimi, is that it does very well on benchmarks, but fails on vibe testing. It feels a little bit over-fitting to the benchmark and less to the use cases.
This was a bad problem with earlier Chinese (Qwen and Kimi K1 in particular) models, but the original DeepSeek delivered and GLM4.6 delivers. They don't diversify training as much as American labs so you'll find more edge cases and the interaction experience isn't quite as smooth, but the models put in work.
Weird, I have gone local for the last 2 years. I use Chinese models 90% of the time, Kimi K2 Thinking, DeepSeekv3.Terminus, Qwen3 and GLM4.6. I'm not vibe testing it but really putting them to use and they do keep up great.
Depends on if they are trained for tool calling, this model is experimentation with new architecture, training methods, etc. It's not designed for tool calling. If you want to tool call, then you should look into DeepSeekv3.1-Terminus.
> DeepSeek-V3.2 introduces significant updates to its chat template compared to prior versions. The primary changes involve a revised format for tool calling and the introduction of a "thinking with tools" capability.
At first, I thought they had gone the route of implementing yet another chat format that can handle more dynamic conversations like that, instead of just using Harmony, but looking at the syntax, doesn't it look exactly like Harmony? That's a good thing, don't get me wrong, but why not mention straight up that they've implemented Harmony, so people can already understand up front that it's compatible with whatever parsing we're using for GPT-OSS?
To push back on naivety I'm sensing here I think it's a little silly to see Chinese Communist Party backed enterprise as somehow magnanimous and without ulterior, very harmful motive.
After using it a couple hours playing around, it is a very solid entry, and very competitive compared with the big US relaeses. I'd say it's better than GLM4.6 and I'm Kimi K2. Looking forward to v4
How will the Google/Anthropic/OpenAI's of the world make money on AI if open models are competitive with their models? What hurt open source in the past was its inability to keep up with the quality and feature depth of closed source competitors, but models seem to be reaching a performance plateau; the top open weight models are generally indistinguishable from the top private models.
Infrastructure owners with access to the cheapest energy will be the long run winners in AI.
> How will the Google/Anthropic/OpenAI's of the world make money on AI if open models are competitive with their models?
So a couple of things. There are going to be a handful of companies in the world with the infrastructure footprint and engineering org capable of running LLMs efficiently and at scale. You are never going to be able to run open models in your own infra in a way that is cost competitive with using their API.
Competition _between_ the largest AI companies _will_ drive API prices to essentially 0 profit margin, but none of those companies will care because they aren't primarily going to make money by selling the LLM API -- your usage of their API just subsidizes their infrastructure costs, and they'll use that infra to build products like chat gpt and claude, etc. Those products are their moat and will be where 90% of their profit comes from.
I am not sure why everyone is so obsessed with "moats" anyway. Why does gmail have so many users? Anybody can build an email app. For the same reason that people stick with gmail, people are going to stick with chatgpt. It's being integrated into every aspect of their lives. The switching costs for people are going to be immense.
They've developed a sparse attention mechanism (which they document and release source code for) to increase model efficiency with long context, as needed for fast & cost-effective extensive RL training for reasoning and agentic use
They've built a "stable & scalable" RL protocol - more capable RL training infrastructure
They've built a pipeline/process to generate synthetic data for reasoning and agentic training
These all combine to build an efficient model with extensive RL post-training for reasoning and agentic use, although they note work is still needed on both the base model (more knowledge) and post-training to match frontier performance.
What version is actually running on chat.deepseek.com?
It refuses to tell me when asked, only that it's been train with data up until July 2024, which would make it quite old. I turned off search and asked it for the winner of the US 2024 election, and it said it didn't know, so I guess that confirms it's not a recent model.
59 comments
[ 3.5 ms ] story [ 96.5 ms ] threadhttps://x.com/deepseek_ai/status/1995452641430651132
The nature of the race may change as yet though, and I am unsure if the devil is in the details, as in very specific edge cases that will work only with frontier models ?
The US labs aren't just selling models, they're selling globally distributed, low-latency infrastructure at massive scale. That's what justifies the valuation gap.
Edit: It looks like Cerebras is offering a very fast GLM 4.6
It reminds me, in an encouraging way, of the way that German military planners regarded the Soviet Union in the lead-up to Operation Barbarossa. The Slavs are an obviously inferior race; their Bolshevism dooms them; we have the will to power; we will succeed. Even now, when you ask questions like what you ask of that era, the answers you get are genuinely not better than "yes, this should have been obvious at the time if you were not completely blinded by ethnic and especially ideological prejudice."
1. Chinese models typically focus on text. US and EU models also bear the cross of handling image, often voice and video. Supporting all those is additional training costs not spent on further reasoning, tying one hand in your back to be more generally useful.
2. The gap seems small, because so many benchmarks get saturated so fast. But towards the top, every 1% increase in benchmarks is significantly better.
On the second point, I worked on a leaderboard that both normalizes scores, and predicts unknown scores to help improve comparisons between models on various criteria: https://metabench.organisons.com/
You can notice that, while Chinese models are quite good, the gap to the top is still significant.
However, the US models are typically much more expensive for inference, and Chinese models do have a niche on the Pareto frontier on cheaper but serviceable models (even though US models also eat up the frontier there).
With caching the expensive US models end up being like 2x the price (e.g sonnet) and often much cheaper (e.g gpt-5 mini)
If they start caching then US companies will be completely out priced.
Frontier models are far exceeding even the most hardcore consumer hobbyist requirements. This is even further
I don't want to make big generalizations. But one thing I noticed with chinese models, especially Kimi, is that it does very well on benchmarks, but fails on vibe testing. It feels a little bit over-fitting to the benchmark and less to the use cases.
I hope it's not the same here.
which version are you?
我是DeepSeek最新版本模型！ 如果你想了解具体的版本号信息，我建议你：
我具备DeepSeek的所有最新功能特性，包括： 如果你需要知道精确的版本号用于技术对接或其他特定用途，最好直接查阅官方技术文档，那里会有最准确和详细的技术规格说明。有什么其他问题我可以帮你解答吗？
At first, I thought they had gone the route of implementing yet another chat format that can handle more dynamic conversations like that, instead of just using Harmony, but looking at the syntax, doesn't it look exactly like Harmony? That's a good thing, don't get me wrong, but why not mention straight up that they've implemented Harmony, so people can already understand up front that it's compatible with whatever parsing we're using for GPT-OSS?
Infrastructure owners with access to the cheapest energy will be the long run winners in AI.
Given how often new models come out, it’s also easier to update an API call than constantly deploying model upgrades.
But in the long run, I hope open source wins out.
So a couple of things. There are going to be a handful of companies in the world with the infrastructure footprint and engineering org capable of running LLMs efficiently and at scale. You are never going to be able to run open models in your own infra in a way that is cost competitive with using their API.
Competition _between_ the largest AI companies _will_ drive API prices to essentially 0 profit margin, but none of those companies will care because they aren't primarily going to make money by selling the LLM API -- your usage of their API just subsidizes their infrastructure costs, and they'll use that infra to build products like chat gpt and claude, etc. Those products are their moat and will be where 90% of their profit comes from.
I am not sure why everyone is so obsessed with "moats" anyway. Why does gmail have so many users? Anybody can build an email app. For the same reason that people stick with gmail, people are going to stick with chatgpt. It's being integrated into every aspect of their lives. The switching costs for people are going to be immense.
They've built a "stable & scalable" RL protocol - more capable RL training infrastructure
They've built a pipeline/process to generate synthetic data for reasoning and agentic training
These all combine to build an efficient model with extensive RL post-training for reasoning and agentic use, although they note work is still needed on both the base model (more knowledge) and post-training to match frontier performance.
"create me a svg of a pelican riding on a bicycle"
https://www.svgviewer.dev/s/FhqYdli5
It refuses to tell me when asked, only that it's been train with data up until July 2024, which would make it quite old. I turned off search and asked it for the winner of the US 2024 election, and it said it didn't know, so I guess that confirms it's not a recent model.