DeepSeek R1
The original R1 uses several training stages for mathematics, code and general reasoning. Its downloadable checkpoint is separate from later R1 updates and smaller distilled models.
Cold-start examples make reasoning easier to read
R1 starts with readable reasoning examples before reinforcement learning. A later supervised stage combines reasoning data with writing and other general tasks. The paper describes a language-consistency reward that improved readability despite a small reasoning-performance tradeoff. These steps distinguish R1 from the direct reinforcement-learning experiment in R1-Zero.DeepSeek-R1 research paper and author record
Active parameters do not describe all stored weights
The checkpoint defines 61 layers and 256 routed experts, with eight selected per token and one shared expert. Its file configuration specifies FP8 block quantization. The stored max_position_embeddings value is 163,840; that configuration value is separate from the publisher's advertised context.Original R1 checkpoint configuration
The old API name no longer selects this checkpoint
The former deepseek-reasoner API moved through newer models and was discontinued on 24 July 2026.DeepSeek dated release history
The original R1 files remain downloadable under their own repository identifier.Original DeepSeek-R1 model card
Specifications & access
| Detail | Value and conditions |
|---|---|
| Launch date | |
| Access | Downloadable |
| Model identifiers | DeepSeek-R1deepseek-ai/DeepSeek-R1 |
| Context window | 131,072 tokens Published 128K context for the weight model. Original DeepSeek-R1 model card |
| Parameters | 671 billion total; 37 billion active per token.Original DeepSeek-R1 model card |
| Base model | DeepSeek-V3-Base.Original DeepSeek-R1 model card |
| Weight license | MIT; commercial use and distillation permitted.DeepSeek announcement of R1 and R1-Zero, 20 January 2025 |
Related models
- DeepSeek R1-0528
DeepSeek · DeepSeek R1
Updated full-size R1 reasoning model with more post-training computation, deeper reasoning and revised prompt handling.
Launch date Downloadable - DeepSeek R1-0528 Qwen3 8B
DeepSeek · DeepSeek R1
Dense reasoning model trained from Qwen3-8B-Base using reasoning examples generated by the larger R1-0528 model.
Launch date May 2025Downloadable - DeepSeek R1-Zero
DeepSeek · DeepSeek R1
Research reasoning model trained with reinforcement learning directly from a pretrained base, without an initial supervised fine-tuning stage.
Launch date Research releaseDownloadable