Model benchmarks
google/gemma-4-31b local LLM performance
As of October 2026, google/gemma-4-31b runs at up to 20.0 tok/s for local inference (best of 5 community benchmark runs across 2 GPUs).
Model size
31B
Peak speed
20.0 tok/s
Average speed
15.9 tok/s
Avg PP
388.7 tok/s
Min memory
37.1 GB
Max context
65,536 tokens
Avg output / run
9,558 tokens
Avg runtime / run
10m 34s
Avg quality
67.1
Benchmark runs
5
GPUs tested
2
Quality by task
Average LLM-judged quality (0–100) with the run-to-run spread shown as a P5–P95 band, overall and for each benchmark task, across all 5 runs. The low and high columns show how much the judge’s score varies between runs, and need at least two runs to display.
| Task | P5 (low) | Avg | P95 (high) |
|---|---|---|---|
| Overall | 42.2 | 67.1 | 77.6 |
| Agent Workflow | 69.4 | 80.7 | 86.0 |
| Code Generation | 11.4 | 54.4 | 72.4 |
| Role Play & Narrative | 72.3 | 76.8 | 81.9 |
| Research & Analysis | 13.5 | 56.4 | 72.5 |
Performance by hardware and tool
Every hardware/tool/quantization combination google/gemma-4-31b has been benchmarked on, ranked by peak token generation speed. Last updated October 2026.
| Hardware | Tool | Quant | Best tok/s | Avg tok/s | Memory | Context | Quality | Runs |
|---|---|---|---|---|---|---|---|---|
| Apple M5 Max | LM Studio | — | 20.0 tok/s | 17.5 tok/s | n/a | 65,536 tokens | 73.6 | 2 |
| Apple M5 Max | LM Studio | Q8_0 | 15.2 tok/s | 15.2 tok/s | 37.1 GB | 65,536 tokens | 76.7 | 2 |
| AMD Radeon RX 7900 XTX | LM Studio | — | 14.2 tok/s | 14.2 tok/s | n/a | 65,536 tokens | 34.8 | 1 |
Benchmark runs
All 5 google/gemma-4-31b runs submitted so far — expand one for its hardware, quality breakdown and per-scenario detail.
Frequently asked questions
- Is google/gemma-4-31b good for coding?
- In our benchmarks, google/gemma-4-31b scores 54.4/100 for coding. It runs at about 15.9 tok/s, so if you want more speed, Huihui-Qwen3.8-27B-abliterated-Swift-1.5-GSQ-RCO-IQ3_S-mtp is faster (~43.5 tok/s) and still scores well for coding (86.6/100).
- Is google/gemma-4-31b good for agentic (tool-using) tasks?
- In our benchmarks, google/gemma-4-31b scores 80.7/100 for agentic workflows. It runs at about 15.9 tok/s, so if you want more speed, granite-4.2-8b-Q4_K_L is faster (~71.6 tok/s) and still scores well for agentic workflows (92.9/100).
- How fast is google/gemma-4-31b for local inference?
- Across 5 community benchmark runs, google/gemma-4-31b reaches up to 20.0 tok/s and averages 15.9 tok/s, with the fastest results on Apple M5 Max.
- How much memory does google/gemma-4-31b need?
- The leanest observed configuration used about 37.1 GB of memory (quantizations tested: Q8_0).
- Which tools have been used to run google/gemma-4-31b?
- Benchmarks were submitted using LM Studio. Results are community-contributed and updated as new runs arrive.