Model comparison
Gemma 4 31B vs Qwen 3.6 35B A3B
Gemma 4 31B vs Qwen 3.6 35B A3B for local inference — VRAM, tokens/sec on RTX 4090 and M4 Max, and quality benchmarks side by side. Qwen 3.6 35B A3B leads on quality, while Qwen 3.6 35B A3B decodes faster on an RTX 4090.
| Metric |  Gemma 4 31B |  Qwen 3.6 35B A3B |
|---|---|---|
| Parameters | 30.7B | 35B |
| Active params (MoE) | dense | 3B |
| Context window | 256K | 262K |
| Quality tier | 86 | 98 |
| MMLU-Pro | 85.2 | 85.2 |
| GPQA Diamond | 84.3 | 86.0 |
| SWE-bench Verified | — | 73.4 |
| LiveCodeBench | 80.0 | 80.4 |
| VRAM (Q4_K_M) | 18.7 GB | 21.3 GB |
| Speed — RTX 4090 (tok/s) | 13.0 | 34.1 |
| Speed — M4 Max (tok/s) | 25.5 | 43.7 |
Which should you run?
Qwen 3.6 35B A3B scores higher on the quality benchmarks, so pick it when answer quality matters most. Qwen 3.6 35B A3B is faster to decode on an RTX 4090 (34 vs 13 tok/s). Gemma 4 31B needs less VRAM (18.7 GB at Q4_K_M), so it fits on smaller GPUs.