Gemma 4 31B
Google · released 2026-04-01 · Apache-2.0 license
Gemma 4 31B is Google's largest dense open model (Apache-2.0, April 2026): 85.2% MMLU-Pro, 89.2% AIME 2026, 80.0% LiveCodeBench, 84.3% GPQA. Targets strong reasoning-per-parameter on a single 80GB GPU.
Key specs
| Type | Local open-weight |
|---|---|
| Parameters | — |
| Architecture | — |
| Context window | — |
| Knowledge cutoff | 2025-01-01 |
| Modalities | text, vision |
| Recommended backends | — |
| Minimum viable rig | — |
Benchmark scores
| GPQA Diamond | 84.3% |
|---|---|
| SWE-bench Verified | — |
| AIME | 89.2% |
| MMLU-Pro | 85.2% |
| BFCL v3 (tool use) | — |
| Composite score | 6.87 |
| Community rating | No reviews yet |
Strengths & weaknesses
Strengths: High intelligence-per-parameter; strong math (AIME 89.2) for a 31B dense model; Multimodal (text+image), 256K context, runs on a single 80GB GPU; Apache-2.0
Weaknesses: No first-party agentic-coding (SWE-bench) scores published; HLE (19.5) below proprietary frontier models