In January 2026, DeepSeek open-sourced its R1 reasoning model directly, and the community's reaction was immediate: no one expected open-source models to have advanced this far.
First, the hard numbers:
- AIME mathematics competition: 77.5 points
- MATH-500: 96.2 points
- Codeforces programming: 94th percentile
- MMLU: 0.849
Taken together, these results put R1 roughly on par with OpenAI's o1. It is a triple threat in mathematics, code, and logical reasoning.
But the distilled versions were even more surprising. R1-Distill-Qwen-32B, a "small" distilled model with 32B parameters, directly outperformed o1-mini on several benchmarks. A 32B-parameter model can match or even surpass a closed-source reasoning model, which is a huge boon for small teams and independent developers — it can run on consumer-grade GPUs.
R1's reasoning approach follows a structured Chain-of-Thought path. In long-context scenarios, it clearly exhibits a "plan → execute → reflect → revise" workflow. The 128K context window provides enough room for the reasoning chain to fully unfold without worrying about mid-process truncation.
From an industry perspective, R1's real impact lies in pricing logic. Previously, the premium on closed-source models largely relied on a "reasoning ability moat" — if you wanted good reasoning, you had to pay. Now R1 has punched a big hole in that wall. What will closed-source vendors use to maintain their pricing advantage? Brand and ecosystem alone?
This question will become increasingly acute over the next few quarters.
Sources: CocoLoop, InfoQ DeepSeek R1 launch coverage, Hugging Face model page