Gemma 4

Four-variant family released under Apache 2.0 (a notable license upgrade from prior Gemma releases). Includes a 31B dense flagship, 26B-A4B MoE (4B active), and edge variants E4B and E2B. 256K token context for the larger models. Multimodal: text, image, video, audio.

Architecture features alternating attention, dual RoPE, Per-Layer Embeddings (PLE), and shared KV cache. Gemma 4 31B achieves 85.2% MMLU Pro, 89.2% AIME 2026, 80.0% LiveCodeBench v6, and Codeforces Elo 2,150. Ranked #3 open model on Arena AI text leaderboard. Crossed 2M downloads shortly after release.

The Gemma 4 Technical Report (arXiv, July 2, 2026; 300+ authors) details the family: dense and MoE architectures from 2.3B to 31B, improved vision and audio encoders at every size, the unified encoder-free architecture of the 12B variant (raw audio and image patches in, no separate encoders), and a thinking mode generating reasoning traces before responding.

June 5, 2026: deployment-ready QAT (Quantization-Aware Training) 4-bit checkpoints shipped across the family on HuggingFace — including GGUF + unquantized + CT variants for E2B, E4B, 12B, 26B-A4B, and 31B (~20 repos total).