Gemma 4 model card (31B instruction-tuned)

Summary

Google's Hugging Face model card for Gemma 4, its open-weight family built on Gemini 3 research and released under Apache 2.0. Google's launch post could not be fetched on 2026-10-04, so the announcement date (3 April 2026) comes from Google's developer forum and news reports. The card cites a technical report dated July 2026 and lists a 12B model that launch coverage did not mention, so it was probably added later.

Key points

• Sizes: E2B (2.3B effective), E4B (4.5B effective), 12B, 26B A4B (MoE, about 3.8B active), and 31B dense.
• Context: 128K tokens for E2B/E4B; 256K for 12B, 26B A4B, and 31B.
• Modalities: text and image for all sizes, and video as frame sequences; audio input on E2B, E4B, and 12B.
• Attention interleaves local sliding-window layers with global layers to keep long contexts cheap.
• Configurable thinking modes. Self-reported scores for 31B: 85.2% on MMLU-Pro, 80.0% on LiveCodeBench v6.
• Licence: Apache 2.0, a change from the custom Gemma licence used for earlier generations.

Why it matters

The strongest small open-weight family from a frontier lab, sized for phones, laptops, and single GPUs, and now under a standard permissive licence.