rStar2-Agent-14B
A 14B math reasoner trained with agentic RL that thinks in Python — matching 671B DeepSeek-R1 on AIME with shorter answers.
About
rStar2-Agent takes the long-chain-of-thought idea one step further: the model doesn't just think longer, it thinks with tools. During reasoning it writes and runs Python code, reads the execution feedback, and revises its approach, the way a person would debug a proof on a whiteboard. Microsoft trained this behavior with agentic reinforcement learning on just 64 AMD MI300X GPUs, finishing in 510 RL steps over a single week.
The result is striking for a 14B model: 80.6% average pass@1 on AIME 2024 and 69.8% on AIME 2025, ahead of the 671B DeepSeek-R1 while producing noticeably shorter responses. The code and training recipes are MIT-licensed on GitHub, though no model weights have been released, so reproducing it means running the RL pipeline yourself.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Paper-reported (arXiv 2508.20722): 80.6% avg pass@1 on AIME24, 69.8% on AIME25, surpassing DeepSeek-R1 (671B) with shorter responses
- Training cost: 510 RL steps on 64 AMD MI300X GPUs in one week
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Scored 8/10 (TIER verified, metrics-v1). Evidence: peer-style technical report (arXiv 2508.20722) from Microsoft Research with clear AIME24/25 numbers vs DeepSeek-R1; MIT-licensed code at github.com/microsoft/rStar (1.4k+ stars); corroborated by independent coverage. Strengths: transparent training recipe, permissive code license, remarkable efficiency story. Docked for: (1) no released model weights — reproducing requires running the RL pipeline; (2) headline numbers are author-reported. License: MIT (code); no weights released.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 8.0/10
- Openness · license + docs
- 7.0/10
- Manual review · sanity read + red-flag check
- 8.0/10
- · Microsoft Research; paper Aug 2025 (arXiv 2508.20722)
- · 14B: 80.6% AIME24 / 69.8% AIME25, beating DeepSeek-R1-671B
- · No released weights; benchmarks author-reported