A.I.
Google · stub
Gemma 4
Gemma 4 is Google DeepMind’s open-weight model family, described as built from Gemini 3 research and technology for intelligence-per-parameter. It is a separate stack from hosted Gemini Flash / Pro API SKUs. DeepMind’s Gemma 4 hub lists sizes including 12B, 26B, and 31B, plus edge-oriented E4B and E2B variants in the performance table. Weights are offered via Hugging Face, Ollama, Kaggle, and LM Studio links on that hub. The Gemini API pricing page also lists Gemma 4 with free-tier access and paid tier marked not available.
- Status
- Current, open weights
- AA score
- 19 (Gemma 4 31B)Artificial Analysis Intelligence Index, one independent score across many tests, fetched 22 Sep 2026
- Last checked
- 22 Sep 2026newest source date on this page
DeepMind’s published table (IT Thinking variants) includes Arena AI (text, as of 2 Apr 2026) Elo 1452 for Gemma 4 31B, MMMLU 85.2%, MMMU Pro 76.9%, AIME 2026 89.2%, LiveCodeBench v6 80.0%, GPQA Diamond 84.3%, and τ2-bench Retail 86.4%. Capability framing on the hub: agentic workflows with native function calling, multimodal reasoning, support for 140 languages, fine-tuning, and local/edge deployment.
Artificial Analysis Intelligence Index (secondary): Gemma 4 31B 19* on the Artificial Analysis leaderboard (fetched 22 Sep 2026). AA also lists other Gemma 4 sizes; those Index figures are on the same leaderboard if needed.
This page is a short stub for the open-weight door. Full six-heading treatment is reserved for hosted Gemini Flash and Pro.