Gemma 4 31B

Gemma 4 31B is the largest dense model in Google’s flagship open series. It natively supports image‑and‑text inputs and reaches over 1,800 tokens/s on Cerebras Cloud. The model is released under Apache 2.0, allowing free commercial use, modification and redistribution.

Its parameter efficiency is remarkable: 31B parameters can match or surpass the performance of earlier 200B+ models on math and coding benchmarks. In direct comparisons, it often uses 5× fewer tokens to produce a clearer, more concise answer than similar‑sized models, saving significant inference time.

The main limitation is the 31B ceiling – it may underperform larger models on complex agentic planning tasks (e.g., SWE‑bench). Google has not disclosed distillation details, so the model’s full potential remains partly opaque. Local deployment requires ≥32 GB VRAM or unified memory; a new desktop build (RTX 5090 32G) costs about ¥42k, while an Apple MacBook Air M5 32GB costs about ¥13k. The quantized model occupies roughly 20 GB.

It is well suited for developers and researchers who need local, low‑latency, high‑quality reasoning, especially for programming assistance and math.