Fastest
Qwen3.6-35B-A3B-Uncensored-Heretic-MLX-4bit
35B
140.2 tok/s
Highest token generation speed at 140.2 tok/s.
Hardware Benchmarks
Best Local LLM Performance for Apple M4 Pro
Compare Apple M4 Pro benchmark results and see which models actually earn the best completed quality scores.
Hardware Overview
GPU-specific specifications for local LLM planning.
- VRAM
- 48GB
Recommended Models
Models we recommend for Apple M4 Pro.
Best quality
Qwen3.8-27B-Uncensored-MLX-6bit
27B
29.2 tok/sQuality 90.3
Highest completed quality score at 90.3.
Largest context
gemma4:26b-a4b-it-qat
25.2B
45.3 tok/sContext 524,288 tokens
Largest context that fits in 48GB VRAM, at 524,288 tokens.
Frequently asked questions
Which models actually run best on the Apple M4 Pro, by task, from community benchmark data.
- What is the best local model for coding on the Apple M4 Pro?
- On the Apple M4 Pro, TensorFold/Qwen3.8-27B-MLX-4bit ranks first for coding in our benchmarks (86.3/100 at ~13.4 tok/s). tensorfold-qwen38-27b is faster (~25.0 tok/s) and still scores well (84.3/100) — a good pick if you'd rather trade a little quality for speed.
- What is the best local model for agentic workflows on the Apple M4 Pro?
- On the Apple M4 Pro, Qwen3.8-27B-Uncensored-MLX-6bit ranks first for agentic workflows in our benchmarks (91.5/100 at ~29.2 tok/s). nerkyor/Qwen3.8-27B-Coder390-EfficientThink-Opus5.5-GPT6Astra-Grok4.7-DSV4Pro-K3-SFT-RLOO-MTP-DFlash2:Q4LynnStyle-MTP is faster (~36.1 tok/s) and still scores well (91.1/100) — a good pick if you'd rather trade a little quality for speed.
Benchmark Results
Completed quality-ranked benchmark results for Apple M4 Pro.