DeepSeek R1: The Release That Made Reasoning Models Downloadable

What DeepSeek R1 is

DeepSeek R1 is the reasoning model that changed what open weights meant. It is a 671 billion parameter mixture-of-experts model with 37 billion activated per token, published alongside DeepSeek-R1-Zero at the same size, and DeepSeek's claim for it was direct: performance comparable to OpenAI o1 across mathematics, code and reasoning. Before it, frontier-level chain-of-thought reasoning was something you rented; after it, it was something you could download. One note on the specification: this site's roster recorded a 64K context length until DeepSeek's own model card was read for this guide, which states 128K; the roster has been corrected to match the first-party source.

The distillations were the larger event

R1 shipped with a set of smaller models distilled from its reasoning traces, and for most people those mattered more than the 671 billion parameter original, because they run on ordinary hardware. The published figures make the case. DeepSeek-R1-Distill-Qwen-32B records 72.6 on AIME 2024 pass@1, 83.3 on cons@64, 94.3 on MATH-500, 62.1 on GPQA Diamond, 57.2 on LiveCodeBench and a CodeForces rating of 1691. DeepSeek-R1-Distill-Llama-70B reaches 70.0, 86.7, 94.5, 65.2 and 57.5 with a CodeForces rating of 1633. Set those against the contemporaries in the same table: GPT-4o-0513 at 9.3 on AIME 2024, Claude 3.5 Sonnet at 16.0, and o1-mini at 63.6. A 32-billion-parameter model you could run locally was outscoring o1-mini on competition mathematics, and that is the result that reset expectations for what small open models could do.

The evaluation conditions matter here

DeepSeek publish how these numbers were produced, and the parameters are not incidental. Maximum generation length was set to 32,768 tokens for every model — a reasoning model given a short budget scores very differently. For benchmarks requiring sampling they used a temperature of 0.6, top-p of 0.95, and generated 64 responses per query to estimate pass@1. That last detail is the one to carry: a pass@1 figure averaged over 64 samples is a much more stable measurement than a single attempt, and it is not the same thing as what you get from one call in production. The cons@64 column in the table is explicitly the consensus over those 64 responses, which is a different and considerably more expensive way to use a model.

LMArena and later releases

On LMArena's text leaderboard DeepSeek R1 has 1398 Elo (rank 168); by category, 1444 Elo in Coding (rank 167), 1412 Elo in Math (rank 144), 1402 Elo in Expert (rank 180) and 1374 Elo in Creative Writing (rank 154). Within DeepSeek's own line, R1-0528 increased reasoning depth, and the V3.1 and V3.2 releases carry larger context windows. All of them are open-weight and self-hostable.

Price and distilled variants

R1-0528 is the direct successor in DeepSeek's reasoning line, with a published accuracy gain. The distilled variants are smaller models with documented evaluation conditions. R1 is priced at roughly not published per million input tokens.

Training pipeline and the distilled checkpoints

DeepSeek-R1-Zero was trained with large-scale reinforcement learning directly on the base model, with no supervised fine-tuning first; DeepSeek report it developed self-verification, reflection and long chains of thought, along with endless repetition, poor readability and language mixing. R1 adds cold-start data before reinforcement learning, in a pipeline of two reinforcement-learning stages and two supervised stages. Both are trained from DeepSeek-V3-Base. Six distilled dense models were released: Qwen-based at 1.5B, 7B, 14B and 32B, and Llama-based at 8B and 70B, the 70B built on Llama-3.3-70B-Instruct. The weights are MIT-licensed. In DeepSeek's table R1 records 90.8 on MMLU, 71.5 on GPQA Diamond, 65.9 on LiveCodeBench and a Codeforces rating of 2029, against o1-1217's 91.8, 75.7, 63.4 and 2061.

Official Resources