Curiosity · AI Model

Google MathGemma

MathGemma is a math-specialised variant of the Gemma open-weights family from Google DeepMind, continuing the RecurrentGemma / CodeGemma / PaliGemma 'specialist Gemma' strategy. Released alongside Gemma 2/3, it is fine-tuned on curated mathematics corpora, MathPile-style datasets, and synthetic Lean / Isabelle proofs, producing stronger step-by-step reasoning on GSM8K, MATH, and olympiad-style benchmarks than the base Gemma models at matching size. Targeted at academic and educational deployments rather than frontier reasoning research.

Model specs

Vendor
Google DeepMind
Family
Gemma (specialist)
Released
2025-03
Context window
8,192 tokens
Modalities
text, code

Strengths

  • Much stronger math performance than base Gemma at matched size
  • Step-by-step reasoning tuned for classroom explanations
  • Open weights under Gemma licence
  • Pairs well with Lean / Isabelle toolchains

Limitations

  • Weaker on non-math tasks than base Gemma
  • Smaller than frontier math models like Minerva and DeepSeek-Math
  • No image/diagram input
  • Licence has Gemma-style usage conditions

Use cases

  • Open-weights math tutoring apps
  • Step-by-step solutions for homework platforms
  • Proof-sketch generation feeding into Lean / Isabelle
  • Education-focused fine-tuning targets

Benchmarks

BenchmarkScoreAs of
GSM8K (math word problems)≈80% (small variant)2025-03
MATH≈45%2025-03

Frequently asked questions

What is MathGemma?

MathGemma is a math-specialised Gemma family fine-tune from Google DeepMind, optimised for step-by-step math reasoning and proof-sketch generation.

How does MathGemma compare to AlphaProof?

AlphaProof is a frontier RL-trained Lean prover; MathGemma is a smaller open fine-tune intended for reasoning and education, not competition-level formal proofs.

Can I fine-tune MathGemma further?

Yes — weights are open under the Gemma licence, and LoRA / full fine-tuning recipes work like any other Gemma checkpoint.

Sources

  1. Google AI — Gemma family — accessed 2026-04-20
  2. MathPile dataset (reference training corpus) — accessed 2026-04-20