M3 Ultra — benchmark record

M3 Ultra local LLM benchmarks across 256 GB, 512 GB RAM tiers on Apple Silicon Mac. 26 published rows across 15 models with explicit evidence state and RAM-tier comparison. Peak published speed is 370.0 tok/s.

26Benchmark rows

15Models tested

2RAM configurations

370.0Fastest 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.

256 GBM3 Ultra (256 GB)13 rows · 10 models370.0 tok/s peak 512 GBM3 Ultra (512 GB)13 rows · 6 models43.0 tok/s peak

All benchmark rows — M3 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
M3 Ultra (256 GB)	Qwen 3 0.6B	4bit	—	370.0 tok/s	—	MLX	ref
M3 Ultra (256 GB)	Qwen3.5-9B	4bit	5.1 GB	106.0 tok/s	—	MLX	ref
M3 Ultra (256 GB)	Qwen3.5-35B-A3B	4bit	19.6 GB	95.0 tok/s	—	MLX	ref
M3 Ultra (256 GB)	Qwen3.5-35B-A3B	8bit	36.9 GB	80.0 tok/s	—	MLX	ref
M3 Ultra (256 GB)	Qwen3-Coder-Next	4bit	44.9 GB	74.0 tok/s	—	MLX	ref
M3 Ultra (256 GB)	Qwen3-Coder-Next	6bit	64.8 GB	66.0 tok/s	—	MLX	ref
M3 Ultra (256 GB)	GLM-4.7-Flash	8bit	—	58.0 tok/s	—	MLX	ref
M3 Ultra (256 GB)	Qwen3.5-122B-A10B	MXFP4	65.0 GB	57.0 tok/s	—	MLX	ref
M3 Ultra (256 GB)	GLM-4.5-Air	4bit	60.3 GB	54.0 tok/s	—	MLX	ref
M3 Ultra (256 GB)	Devstral Small 2 24B	4bit	13.4 GB	47.0 tok/s	—	MLX	ref
M3 Ultra (256 GB)	Qwen3.5-122B-A10B	8bit	129.8 GB	43.0 tok/s	—	MLX	ref
M3 Ultra (512 GB)	Devstral Small 1.1	4bit	—	43.0 tok/s	—	LM Studio	ref
M3 Ultra (512 GB)	Qwen3.5-397B-A17B	q4.1bit	—	40.2 tok/s	—	MLX	ref
M3 Ultra (256 GB)	Qwen3.5-27B	4bit	15.3 GB	38.0 tok/s	—	MLX	ref
M3 Ultra (512 GB)	Qwen 3 235B-A22B	4bit	—	27.4 tok/s	—	LM Studio	ref
M3 Ultra (256 GB)	Qwen 3 235B-A22B	Q4	—	27.0 tok/s	—	MLX	ref
M3 Ultra (512 GB)	GLM-5	4bit	391.8 GB	16.7 tok/s	187.0 tok/s	MLX	ref
M3 Ultra (512 GB)	Llama 3.3 70B	4bit	—	15.5 tok/s	150.0 tok/s	LM Studio	ref
M3 Ultra (512 GB)	GLM-5	4bit	394.1 GB	13.7 tok/s	180.1 tok/s	MLX	ref
M3 Ultra (512 GB)	GLM-5	4bit	396.7 GB	13.2 tok/s	154.1 tok/s	MLX	ref
M3 Ultra (512 GB)	GLM-5	4bit	402.7 GB	12.0 tok/s	117.4 tok/s	MLX	ref
M3 Ultra (512 GB)	Gemma 3 27B	bf16	52.6 GB	11.2 tok/s	—	LM Studio	ref
M3 Ultra (512 GB)	GLM-5	4bit	415.4 GB	10.7 tok/s	77.7 tok/s	MLX	ref
M3 Ultra (512 GB)	Llama 3.3 70B	4bit	—	9.6 tok/s	103.0 tok/s	LM Studio	ref
M3 Ultra (512 GB)	Llama 3.3 70B	8bit	—	8.5 tok/s	150.0 tok/s	LM Studio	ref
M3 Ultra (512 GB)	Llama 3.3 70B	8bit	—	6.5 tok/s	101.0 tok/s	LM Studio	ref

Models tested on M3 Ultra

Devstral Small 1.1 Devstral Small 2 24B Gemma 3 27B GLM-4.5-Air GLM-4.7-Flash GLM-5 Llama 3.3 70B Qwen 3 0.6B Qwen 3 235B-A22B Qwen3-Coder-Next Qwen3.5-9B Qwen3.5-27B Qwen3.5-35B-A3B Qwen3.5-122B-A10B Qwen3.5-397B-A17B

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 →