M4 Ultra — benchmark record

M4 Ultra local LLM benchmarks across 192 GB RAM tiers on Apple Silicon Mac. 8 published rows across 8 models with explicit evidence state and RAM-tier comparison. Peak published speed is 45.0 tok/s.

8Benchmark rows

8Models tested

1RAM configurations

45.0Fastest avg tok/s

RAM configurations

Each configuration differs only in unified memory. More RAM = larger models fit. Throughput is similar across RAM tiers at the same model size.

192 GBM4 Ultra (192 GB)8 rows · 8 models45.0 tok/s peak

All benchmark rows — M4 Ultra

Sorted by avg tok/s descending. Click source badge to see original measurement.

Chip (RAM)	Model	Quant	RAM req.	Avg tok/s	Prompt tok/s	Runtime	Source
M4 Ultra (192 GB)	Mistral Small 4 119B	Q4_K - Medium	—	45.0 tok/s	—	MLX	ref
M4 Ultra (192 GB)	Qwen 3 32B	Q4_K - Medium	—	32.0 tok/s	—	Ollama	ref
M4 Ultra (192 GB)	Llama 4 Scout 17B-16E	Q4_K - Medium	—	30.0 tok/s	—	MLX	ref
M4 Ultra (192 GB)	Qwen 3 235B-A22B	Q4_K - Medium	—	22.0 tok/s	—	MLX	ref
M4 Ultra (192 GB)	Llama 3.3 70B	Q4_K - Medium	—	18.0 tok/s	—	MLX	ref
M4 Ultra (192 GB)	DeepSeek R1 Distill Llama 70B	Q4_K - Medium	—	16.0 tok/s	—	MLX	ref
M4 Ultra (192 GB)	Qwen 2.5 72B	Q4_K - Medium	—	15.0 tok/s	—	Ollama	ref
M4 Ultra (192 GB)	gpt-oss 120B	Q4_K - Medium	—	10.0 tok/s	—	MLX	ref

Models tested on M4 Ultra

DeepSeek R1 Distill Llama 70B gpt-oss 120B Llama 3.3 70B Llama 4 Scout 17B-16E Mistral Small 4 119B Qwen 2.5 72B Qwen 3 32B Qwen 3 235B-A22B

Data

benchmarks.json — full dataset · chips.json — chip summaries · benchmarks.csv — CSV export

Data from in-house lab measurements plus community-published benchmarks. See all chip families →