Model benchmarks
gemma-4-31B-it-MLX-8bit local LLM performance
As of September 2026, gemma-4-31B-it-MLX-8bit runs at up to 16.6 tok/s for local inference (best of 3 community benchmark runs across 2 GPUs).
Model size
31B
Peak speed
16.6 tok/s
Average speed
15.5 tok/s
Min memory
35.7 GB
Max context
262,144 tokens
Avg output / run
9,550 tokens
Avg runtime / run
7m 2s
Avg quality
73.3
Benchmark runs
3
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 3 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 | 71.2 | 73.3 | 74.9 |
| Agent Workflow | 74.8 | 79.1 | 81.8 |
| Code Generation | 59.0 | 61.0 | 62.3 |
| Role Play & Narrative | 80.4 | 81.9 | 83.0 |
| Research & Analysis | 68.7 | 71.4 | 73.1 |
Performance by hardware and tool
Every hardware/tool/quantization combination gemma-4-31B-it-MLX-8bit has been benchmarked on, ranked by peak token generation speed. Last updated September 2026.
| Hardware | Tool | Quant | Best tok/s | Avg tok/s | Memory | Context | Quality | Runs |
|---|---|---|---|---|---|---|---|---|
| Apple M2 Ultra | oMLX | — | 16.6 tok/s | 16.6 tok/s | 81.3 GB | 262,144 tokens | 74.1 | 1 |
| Apple M5 Max | oMLX | — | 14.9 tok/s | 14.9 tok/s | 35.7 GB | 65,536 tokens | 72.9 | 2 |
Benchmark runs
All 3 gemma-4-31B-it-MLX-8bit runs submitted so far — expand one for its hardware, quality breakdown and per-scenario detail.
Frequently asked questions
- Is gemma-4-31B-it-MLX-8bit good for coding?
- In our benchmarks, gemma-4-31B-it-MLX-8bit scores 61.0/100 for coding. It runs at about 15.5 tok/s, so if you want more speed, IsValorum/Nex-N2.5-mini-APEX-I-MiniPlus-V2.1-GGUF:1 is faster (~25.7 tok/s) and still scores well for coding (85.0/100).
- Is gemma-4-31B-it-MLX-8bit good for agentic (tool-using) tasks?
- In our benchmarks, gemma-4-31B-it-MLX-8bit scores 79.1/100 for agentic workflows. It runs at about 15.5 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 gemma-4-31B-it-MLX-8bit for local inference?
- Across 3 community benchmark runs, gemma-4-31B-it-MLX-8bit reaches up to 16.6 tok/s and averages 15.5 tok/s, with the fastest results on Apple M2 Ultra.
- How much memory does gemma-4-31B-it-MLX-8bit need?
- The leanest observed configuration used about 35.7 GB of memory.
- Which tools have been used to run gemma-4-31B-it-MLX-8bit?
- Benchmarks were submitted using oMLX. Results are community-contributed and updated as new runs arrive.