DeepSeek's new R1-0528 open-source LLM reportedly rivals OpenAI's o3 in coding & reasoning. Features a long context window & improved long-text accuracy.
DeepSeek has quietly rolled out R1-0528, a new version of their open-source LLM, and it’s already making some noise. Interestingly, DeepSeek describes this as just a "minor update" to their R1 model – classic DeepSeek modesty, given what it can do. It’s competing closely with OpenAI's o3 models on coding benchmarks on LiveCodeBench.
Key upgrades include strong coding and reasoning skills, a massive long context window perfect for large documents or codebase, and better accuracy on long texts. It’s also designed to be more careful about hallucinations.
For those already using DeepSeek official API for development, the existing API interfaces remain unchanged. The model is now on Hugging Face, with a free API available via OpenRouter.
This is definitely an interesting development for anyone needing powerful, open-source LLM alternatives. And if this is just the "minor update" or appetizer before a potential R2, well, it already feels like a full main course to the world!
Replies
Hi everyone!
DeepSeek has quietly rolled out R1-0528, a new version of their open-source LLM, and it’s already making some noise. Interestingly, DeepSeek describes this as just a "minor update" to their R1 model – classic DeepSeek modesty, given what it can do. It’s competing closely with OpenAI's o3 models on coding benchmarks on LiveCodeBench.
Key upgrades include strong coding and reasoning skills, a massive long context window perfect for large documents or codebase, and better accuracy on long texts. It’s also designed to be more careful about hallucinations.
For those already using DeepSeek official API for development, the existing API interfaces remain unchanged. The model is now on Hugging Face, with a free API available via OpenRouter.
This is definitely an interesting development for anyone needing powerful, open-source LLM alternatives. And if this is just the "minor update" or appetizer before a potential R2, well, it already feels like a full main course to the world!
Looks like the new DeepSeek has caught up on claude opus no think, although it's not on the leaderboard yet
(Deepseek R1.1 scored the same as claude-opus-4-nothink 70.7% on aider polyglot. Old R1 was 56.9%)
For teams working with large-scale documents or code, have you observed any tangible benefits with R1-0528 compared to other open-source LLMs?
This looks really promising. I love Opus but it’s really expensive. If it can be almost as good as Claude 4 for cheaper, I’m all for it.