gemma-4-31B-it

Gemma 4 31B IT is Google DeepMind's instruction-tuned, open-weight dense multimodal language model released on April 2, 2026 under the Apache 2.0 license. With 30.7 billion parameters, it supports text and image inputs, a 256K-token context window, native system prompts, configurable thinking mode, and function calling for agentic workflows. Architecturally it uses a hybrid attention design interleaving 1024-token sliding-window layers with periodic global attention layers, unified Keys/Values, and Proportional RoPE (p-RoPE) for long-context stability, plus a ~550M-parameter vision encoder.

On official Google benchmarks, Gemma 4 31B IT scores 85.2% on MMLU Pro, 89.2% on AIME 2026 (no tools), 80.0% on LiveCodeBench v6, 84.3% on GPQA Diamond, 76.9% on Tau2 agentic tasks, and 76.9% on MMMU Pro for vision. Independent local testing (April 2026, Ollama on NVIDIA DGX Spark) reported ~85% on both HumanEval and GSM8K at Q8 precision. On LMSYS Chatbot Arena it holds an Elo of ~1451, ranking among the top open models globally and #1 among US open-weight models at launch.

The model targets developers seeking frontier-level reasoning, coding, and multimodal understanding on accessible hardware. Unquantized BF16 fits a single 80GB GPU; quantized variants (Q4_0, Q8, GGUF) enable deployment on consumer workstations. It competes with models many times larger while trailing specialized agentic systems on tool-use benchmarks like MCP-Atlas. Available via Hugging Face, Google AI Studio, Vertex AI, and NVIDIA NIM.