Google Gemma 4 31B

Gemma 4 31B is Google DeepMind's flagship open-source model, released on April 2, 2026, under a genuine Apache 2.0 license with no usage restrictions. Built with the same underlying technology as Gemini 3, it packs 307B total parameters into a dense architecture — not a mixture-of-experts — meaning all parameters participate in every forward pass. This is a deliberate architectural choice that maximizes reasoning quality per parameter at the cost of higher compute per token.

The attention mechanism alternates between local sliding-window attention (1024-token windows) and global attention layers. This hybrid approach gives the model efficient processing of nearby context while maintaining the ability to attend to information anywhere in its 256K token context window. The result is a model that handles both short conversational exchanges and long document analysis without the quality degradation that purely local attention models suffer on longer contexts.

The performance jump from Gemma 3 is staggering: AIME 2026 went from 20.8% to 89.2% — a 4.3x improvement in competition-level math reasoning. It scores 84.3% on GPQA Diamond, 85.2% on MMLU Pro, and 80% on LiveCodeBench v6, putting it in the same tier as closed-source models that cost orders of magnitude more to run. At Q4_K_M quantization (recommended by Unsloth), the model fits in approximately 20GB of VRAM, making it runnable on a single RTX 4090 or equivalent.

Day-1 ecosystem support was comprehensive: llama.cpp, Ollama, LM Studio, vLLM, MLX, and Hugging Face Transformers all had working implementations at launch, with GGUF quantizations from Unsloth available immediately. This breadth of tooling support, combined with the truly permissive license, makes Gemma 4 31B the strongest open-source model available for both local deployment and commercial applications.