Model benchmarks
qwen3.5-4b-mlx local LLM performance
As of November 2026, qwen3.5-4b-mlx runs at up to 84.2 tok/s for local inference (best of 4 community benchmark runs across 3 GPUs).
Model size
4B
Peak speed
84.2 tok/s
Average speed
67.0 tok/s
Avg PP
1886.8 tok/s
Min memory
4.6 GB
Max context
65,536 tokens
Avg output / run
15,596 tokens
Avg runtime / run
5m 8s
Avg quality
57.9
Benchmark runs
4
GPUs tested
3
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 4 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 | 47.7 | 57.9 | 68.4 |
| Agent Workflow | 45.6 | 59.9 | 68.7 |
| Code Generation | 19.9 | 38.1 | 61.0 |
| Role Play & Narrative | 63.1 | 70.2 | 75.1 |
| Research & Analysis | 52.2 | 63.2 | 69.9 |
Performance by hardware and tool
Every hardware/tool/quantization combination qwen3.5-4b-mlx 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 M1 Max | LM Studio | 4bit | 84.2 tok/s | 81.4 tok/s | 4.6 GB | 65,536 tokens | 57.6 | 2 |
| Apple M5 Pro | Ollama | nvfp4 | 71.9 tok/s | 71.9 tok/s | 5.0 GB | 32,768 tokens | 46.6 | 1 |
| Apple M1 Pro | LM Studio | 8bit | 33.4 tok/s | 33.4 tok/s | 8.4 GB | 32,768 tokens | 69.7 | 1 |
Benchmark runs
All 4 qwen3.5-4b-mlx runs submitted so far — expand one for its hardware, quality breakdown and per-scenario detail.
Frequently asked questions
- Is qwen3.5-4b-mlx good for coding?
- In our benchmarks, qwen3.5-4b-mlx scores 38.1/100 for coding. It runs at about 67.0 tok/s, so if you want more speed, ornith-ai/Ornith-1.5-35B-A3B-GGUF is faster (~80.3 tok/s) and still scores well for coding (82.1/100).
- Is qwen3.5-4b-mlx good for agentic (tool-using) tasks?
- In our benchmarks, qwen3.5-4b-mlx scores 59.9/100 for agentic workflows. It runs at about 67.0 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 qwen3.5-4b-mlx for local inference?
- Across 4 community benchmark runs, qwen3.5-4b-mlx reaches up to 84.2 tok/s and averages 67.0 tok/s, with the fastest results on Apple M1 Max.
- How much memory does qwen3.5-4b-mlx need?
- The leanest observed configuration used about 4.6 GB of memory (quantizations tested: 4bit, 8bit, nvfp4).
- Which tools have been used to run qwen3.5-4b-mlx?
- Benchmarks were submitted using LM Studio, Ollama. Results are community-contributed and updated as new runs arrive.