Model benchmarks
unsloth/gemma-4-31B-it-qat-GGUF local LLM performance
As of November 2026, unsloth/gemma-4-31B-it-qat-GGUF runs at up to 50.8 tok/s for local inference (best of 6 community benchmark runs across 1 GPU).
Model size
31B
Peak speed
50.8 tok/s
Average speed
45.5 tok/s
Avg PP
483.3 tok/s
Min memory
16.2 GB
Max context
131,072 tokens
Avg output / run
8,920 tokens
Avg runtime / run
3m 29s
Avg quality
74.4
Benchmark runs
6
GPUs tested
1
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 6 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 | 72.4 | 74.4 | 76.8 |
| Agent Workflow | 78.4 | 81.7 | 84.4 |
| Code Generation | 66.1 | 69.3 | 72.4 |
| Role Play & Narrative | 72.6 | 79.5 | 87.7 |
| Research & Analysis | 57.8 | 67.4 | 74.2 |
Performance by hardware and tool
Every hardware/tool/quantization combination unsloth/gemma-4-31B-it-qat-GGUF has been benchmarked on, ranked by peak token generation speed. Last updated November 2026.
| Hardware | Tool | Quant | Best tok/s | Avg tok/s | Memory | Context | Quality | Runs |
|---|---|---|---|---|---|---|---|---|
| Apple M5 Max | Unsloth Studio | UD-Q4_K_XL | 50.8 tok/s | 48.9 tok/s | 16.2 GB | 131,072 tokens | 74.4 | 5 |
| Apple M5 Max | llama.cpp | — | 28.8 tok/s | 28.8 tok/s | n/a | 65,536 tokens | 74.6 | 1 |
Benchmark runs
All 6 unsloth/gemma-4-31B-it-qat-GGUF runs submitted so far — expand one for its hardware, quality breakdown and per-scenario detail.
Frequently asked questions
- Is unsloth/gemma-4-31B-it-qat-GGUF good for coding?
- In our benchmarks, unsloth/gemma-4-31B-it-qat-GGUF scores 69.3/100 for coding. It runs at about 45.5 tok/s, so if you want more speed, qwen3.8-27b-Swift1.5-Uncensored-ATX-AI3 is faster (~85.5 tok/s) and still scores well for coding (84.8/100).
- Is unsloth/gemma-4-31B-it-qat-GGUF good for agentic (tool-using) tasks?
- In our benchmarks, unsloth/gemma-4-31B-it-qat-GGUF scores 81.7/100 for agentic workflows. It runs at about 45.5 tok/s, so if you want more speed, Nex-N2.5-mini-APEX-Mini is faster (~88.2 tok/s) and still scores well for agentic workflows (91.3/100).
- How fast is unsloth/gemma-4-31B-it-qat-GGUF for local inference?
- Across 6 community benchmark runs, unsloth/gemma-4-31B-it-qat-GGUF reaches up to 50.8 tok/s and averages 45.5 tok/s, with the fastest results on Apple M5 Max.
- How much memory does unsloth/gemma-4-31B-it-qat-GGUF need?
- The leanest observed configuration used about 16.2 GB of memory (quantizations tested: UD-Q4_K_XL).
- Which tools have been used to run unsloth/gemma-4-31B-it-qat-GGUF?
- Benchmarks were submitted using Unsloth Studio, llama.cpp. Results are community-contributed and updated as new runs arrive.