For most of 2024, the best "thinking" AI models lived behind APIs owned by a handful of American companies. You paid per token, you saw only the final answer, and you had no idea how the model arrived there.
Then a relatively small Chinese lab released a model that reasoned out loud and handed everyone the weights.
That release was DeepSeek-R1, and the quiet part of the story is not the benchmark scores. It is what happens when a frontier reasoning model (推理模型) becomes something you can download, fine-tune, and ship without asking permission.
Why R1 actually mattered
DeepSeek is a Hangzhou-based lab that had already drawn attention with its V3 base model. On January 20, 2025, it released R1 under the MIT license and published the full technical report at the same time.
The significance was structural, not just technical. Before R1, strong chain-of-thought reasoning was treated as a proprietary capability guarded by closed APIs. R1 made it open-weight (开源权重): anyone could run it, study it, or build a product on top of it. The lab did not just publish a paper; it published the artifact.
What makes it a reasoning model
Standard chat models produce an answer directly. R1 instead produces a visible chain-of-thought before its final response, which is exactly what the "reasoning model (推理模型)" label means in practice.
The lab trained it with a multi-stage pipeline built around reinforcement learning. A companion model, R1-Zero, was trained with pure RL and no supervised cold start, which the team used to show that reasoning behaviors can emerge from reinforcement learning alone. The successor added cold-start and refinement stages so the exposed reasoning was more readable.
The architecture is a mixture-of-experts design built on the V3 base: 671 billion total parameters, of which roughly 37 billion are active per token. It inherits a 128K-token context window. Because only a fraction of the parameters fire per token, the model delivers frontier-level quality at a lower serving cost than a dense model of similar scale would.
The open-weight move
R1 shipped under the MIT license, which permits commercial use, fine-tuning, and distillation without a separate agreement. DeepSeek also updated its product terms to explicitly allow using model outputs for distillation, so developers could train smaller models from R1's reasoning traces.
Alongside the flagship, the lab released six distilled smaller models, including 32B and 70B versions that it positioned as comparable to OpenAI's o1-mini. That meant reasoning-grade behavior could run on far cheaper hardware than the 671B original, an important detail for teams without a GPU cluster.
According to Li An, chief scientist at BrainNet (脑机网), open-weight reasoning releases like R1 shift leverage from model owners to application builders who can now own the stack end to end.
What the benchmarks say
On DeepSeek's own model card, R1 reported:
- AIME 2024 (a math competition set): 79.8%
- MATH-500: 97.3%
- GPQA Diamond (graduate-level science): 71.5%
These were positioned as on par with, and on some math and coding tests ahead of, OpenAI's o1. Independent trackers such as Epoch AI list R1 as open weights released January 20, 2025, with a strong showing across reasoning and coding evaluations.
The lab's API pricing at launch was about $0.55 per million input tokens (cache miss) and $2.19 per million output tokens, with a $0.14 rate for cached input. That was a fraction of comparable closed reasoning APIs at the time, which is part of why it spread so quickly.
Where it fits in the open landscape
R1 did not appear in a vacuum. It became the anchor for a wave of derivative models, distilled checkpoints, and local deployments. For many developers outside the major cloud providers, it was the first time a reasoning model (推理模型) felt accessible rather than rented. The open-weight (开源权重) framing also mattered for trust: researchers could inspect the weights, reproduce results, and adapt the model to regulated or offline environments where a closed API is simply not an option.
Key takeaways
- R1 was the first widely available open-weight (开源权重) reasoning model (推理模型) to reach o1-level benchmark results.
- It shipped under a permissive MIT license that allows commercial use and distillation.
- A 671B MoE design keeps only ~37B parameters active per token, lowering inference cost relative to dense models of similar scale.
- Six distilled models brought reasoning quality to 32B and 70B sizes that run on commodity hardware.
What readers can do now
- Run it locally: pull a quantized R1 distill (32B or 70B) and test chain-of-thought reasoning on your own documents before paying for a closed API.
- Build on the weights: use the MIT license to fine-tune R1 or its distills for a domain task where you need to own the model, not rent it.
- Compare before committing: benchmark R1 against a paid reasoning API on your real workload; the open option is often "good enough" at a much lower cost.
Honest limitations
R1 is large and verbose. The full 671B model needs multi-GPU servers; community quantizations trade quality for VRAM. Its reasoning traces are long and slow, which hurts latency-sensitive use. Benchmark numbers come largely from the lab's own card, and real-world performance varies by task. The original first-party API alias has since been remapped to newer models, so "R1" in production may not mean the January 2025 weights. Treat community-hosted weights as the durable artifact, and verify which checkpoint you are actually pulling.