DeepSeek R1
The open reasoning model that started the RL reasoning wave — 671B MoE under MIT
About
DeepSeek R1, released January 20, 2025 (refreshed as R1-0528 in May 2025), is the open-weights reasoning model trained with large-scale reinforcement learning to produce long chain-of-thought before answering — the release that proved open models could match closed reasoning systems and reset industry cost expectations. It is a 671B-parameter MoE with 37B active per token, a 128K context window, and six distilled variants (1.5B–70B). Weights are MIT-licensed on Hugging Face, explicitly permitting distillation into other models.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- AIME 2024 79.8%, MATH-500 97.3%, Codeforces ~2029 — provider-reported (Jan 2025 launch)
- R1-0528: AIME 2024 91.4%, AIME 2025 87.5%, ~45–50% fewer hallucinations — provider-reported
- GPQA Diamond 81.0–81.3%, LiveCodeBench 73.3%, SWE-bench Verified 57.6%, Humanity's Last Exam 17.7% — vendor-reported figures aggregated by a third party (not independently verified)
- Artificial Analysis Intelligence Index: 13.1; Math index 76 — independent
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Release date (2025-01-20), MIT license, architecture and pricing verified across many independent sources; AA independent index available. Minor caveats: a 21-month-old model now superseded in DeepSeek's own line by V4 for coding; text-only; Chinese origin raises data-sovereignty questions for some enterprises. Metrics-v1 score 8/10 — adoption 9.4, maintenance 5, openness 10, manual 9.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 9.4/10
- Maintenance · recency, releases, issue load
- 5.0/10
- Openness · license + docs
- 10.0/10
- Manual review · sanity read + red-flag check
- 9.0/10
- · HF top match: deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B
- · HF downloads 1,270,002 (peer-normalized)
- · released 631 days ago
- · manual review: 21 months old, superseded by V4 for coding; data-sovereignty notes for some enterprises