DeepSeek-R1
DeepSeek · LLM model
DeepSeek-R1 is the open reasoning model from January 2025 that brought DeepSeek global attention. It is now legacy: its API endpoint was replaced by the V3.x and then V4 models.
Legacy model Open weights
ReleasedJan 20, 2025
Context window128K tokens
Input → outputtext → text
API price—
Consumer access—
Best for
Self-hosted reasoning research where a well-documented MIT-licensed open model is needed.
Overview
DeepSeek-R1 is the open reasoning model from January 2025 that brought DeepSeek global attention. It is now legacy: its API endpoint was replaced by the V3.x and then V4 models.
Strengths and limits
- Landmark open reasoning model, MIT license
- Widely supported by inference tools and distilled variants
- Strong math scores for its time (MATH-500 97.3)
- No longer served on DeepSeek's API (deepseek-reasoner retired 2026-07-24)
- Far behind the V4 family on coding and agent tasks
Published benchmark scores
AIME 2024 (Pass@1)79.8
Source: huggingface.co
MATH-500 (Pass@1)97.3
Source: huggingface.co
GPQA Diamond (Pass@1)71.5
Source: huggingface.co
SWE Verified (Resolved)49.2
Source: huggingface.co
Scores are as published by the source; settings (reasoning effort, tools, attempts) differ between labs, so compare with care. How to read benchmarks →