On January 20, 2025, a comparatively small Chinese lab released a model that made the AI world reassess what open weights could do. DeepSeek-R1, from DeepSeek (深度求索), was the first openly available model to reach reasoning quality the lab described as on par with OpenAI's o1 — and it shipped under the permissive MIT license, free to download, fine-tune, and run.
The release did more than add a strong model to the shelf. It showed that reasoning could be coaxed out of a model largely through reinforcement learning, and it did so at a price a fraction of closed alternatives.
Key takeaways
- DeepSeek-R1 is a 671-billion-parameter MoE model with 37 billion active per token, built on DeepSeek-V3-Base and released as open weights (开源权重) under the MIT license. It was the first openly available large model (大模型) to reach o1-class reasoning.
- It is a reasoning model: it produces a visible chain-of-thought, trained in part with large-scale reinforcement learning (GRPO) rather than only supervised fine-tuning.
- On launch benchmarks it matched or beat o1 on MATH-500 (97.3%), AIME 2024 (79.8%), and GPQA-Diamond (71.5%), with MMLU at 90.8%.
- DeepSeek distilled R1 into six smaller dense models (1.5B–70B) on Qwen and Llama bases, bringing reasoning to single GPUs.
How a reasoning model is made
Standard chat models answer immediately. Reasoning models deliberate first. DeepSeek-R1 exposes an explicit chain-of-thought before its final answer, trading latency for accuracy on math, code, and science problems.
The training pipeline is the interesting part. DeepSeek started from V3-Base and used a multi-stage process: a reinforcement-learning cold start, two RL stages that discover and refine reasoning behavior, and two supervised fine-tuning. A companion model, R1-Zero, was trained with pure RL and no supervised cold start at all — used to demonstrate that reasoning can emerge from RL alone. The companion paper, "DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning" (arXiv 2501.12948), later appeared in Nature.
The RL method uses verifiable rewards — math correctness, code execution, format checks — rather than a learned reward model. That choice, and the GRPO algorithm, became a template the open community copied widely.
The numbers that got attention
DeepSeek reported scores that placed R1 alongside or above o1 on several hard sets:
- AIME 2024: 79.8% (pass@1).
- MATH-500: 97.3%.
- GPQA-Diamond: 71.5%.
- MMLU: 90.8%.
- Codeforces: 96.3rd percentile.
- SWE-bench Verified: 49.2% resolved.
- LiveCodeBench: 65.9%.
On the launch API, pricing sat around US$0.55 per million input tokens and US$2.19 per million output tokens — far below o1, which at the time cost roughly US$15 per million input and US$60 per million output. That gap, more than any single benchmark, is what rattled markets: a frontier-class reasoning model was available at a small fraction of the incumbent price.
Distillation to everyday hardware
Alongside the flagship, DeepSeek released six distilled models built on Qwen 2.5 and Llama 3 in sizes from 1.5B to 70B. These kept much of R1's reasoning behavior while running on a single consumer or datacenter GPU. The 70B distilled variant, in particular, became a common local-choice for developers who wanted R1-grade reasoning without a multi-GPU cluster.
The open release triggered a wave of derivatives — more than 500 models on Hugging Face within days — and a wave of reproductions (TinyZero, Open-R1, SimpleRL) that tried to redo the recipe.
Why it mattered
Before R1, strong reasoning was treated as a closed-model advantage. R1 moved the bar: open weights could reason, could be self-hosted, and could be had cheaply. The app briefly topped the U.S. App Store, and the cost disclosure fed directly into a broader debate about how much compute frontier training actually requires.
The paper's method drew as much attention as the model. By showing that a pure-RL model (R1-Zero) could develop coherent reasoning with no human-labeled chain-of-thought, DeepSeek suggested that data, not just parameters, was the bottleneck — and that verifiable rewards could replace expensive human annotations at scale. That idea spread fast: within weeks, independent labs had reproduced the recipe on smaller models, and the term "reasoning model" shifted from a closed-lab label to something any team could attempt. The release also pushed other Chinese labs to open their own reasoning weights sooner than planned.
Honest limitations
- R1 is text-only; it has no vision or audio.
- Its SWE-bench Verified score (49.2%) trails later coding-focused models, including some open ones released afterward.
- The prominent benchmark figures are DeepSeek's own, measured under its harness and attempt counts; treat them as directional, not absolute.
- The original first-party API endpoint has since been remapped to newer DeepSeek models; R1 weights remain on Hugging Face and third-party hosts, but the "R1" API label no longer always points to this exact checkpoint.
What readers can do now
- Run a distilled R1 (the 7B or 8B Qwen-based version) locally with Ollama or llama.cpp to see open reasoning on one GPU, free of server costs.
- If you build math or code tooling, benchmark R1 against a closed reasoning model on your own tasks before paying the premium.
- Read the GRPO paper if you train models — the verifiable-reward recipe is the part most reused across the open ecosystem.