rStar-Math
The self-evolved deep-thinking recipe that took a 7B SLM from 58.8% to 90% on MATH — no distillation from giants.
About
rStar-Math is less a model than a recipe, and it changed how people think about small models. Instead of distilling a giant teacher, Microsoft Research Asia trained a 7B policy model and a process reward model from scratch, then let them improve each other over four rounds of self-evolution: MCTS rollouts generate step-by-step verified trajectories, the reward model learns which steps are good, and the policy gets better at searching.
The numbers made the case: Qwen2.5-Math-7B went from 58.8% to 90.0% on MATH, beating o1-preview, and even the 3.8B Phi3-mini reached 86.4%. AIME hit 53.3%, good enough for the top fifth of human contestants. The code and data are MIT-licensed, though there is no single downloadable checkpoint — you run the pipeline to get the model.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
- Paper-reported (arXiv 2501.04519): Qwen2.5-Math-7B 58.8% → 90.0% on MATH (+4.5pp over o1-preview); Phi3-mini 41.4% → 86.4%
- Same paper: 53.3% on AIME (8/15), ranking among the top 20% of human contestants
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (TIER verified, metrics-v1). Evidence: paper (arXiv 2501.04519) from Microsoft Research Asia with Peking University, USTC and Tsinghua; MIT-licensed code at github.com/microsoft/rStar; widely covered (VentureBeat) and highly influential on the SLM reasoning wave. Strengths: landmark efficiency result, permissive license, transparent recipe. Docked for: (1) no released checkpoints — it's a training recipe, not a downloadable model; (2) headline numbers are author-reported. License: MIT (code/data); no weights released.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 6.0/10
- Maintenance · recency, releases, issue load
- 6.0/10
- Openness · license + docs
- 7.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · Microsoft Research Asia; paper 2025-01-08
- · Qwen2.5-Math-7B 58.8% to 90.0% on MATH via self-evolution
- · Training recipe, not a downloadable model; numbers author-reported