M3 Ultra (512 GB) — benchmark record

M3 Ultra (512 GB) local LLM benchmarks on Apple Silicon Mac with 13 published rows across 6 models. Peak published speed is 43.0 tok/s on Devstral Small 1.1. Largest published fit uses 415.4 GB. Published runtimes include LM Studio, MLX.

13Benchmark rows

6Models tested

43.0Fastest avg tok/s (Devstral Small 1.1)

0Lab benchmarks

Quick take

Best published speed here is 43.0 tok/s on Devstral Small 1.1 at 4bit. Largest published memory footprint is GLM-5 at 4bit, using 415.4 GB. Longest published context on this page is 128k. This page is evidence, not the full buying answer.

Based on 13 external benchmarks; no lab runs yet.

Published runtimes: LM Studio, MLX.

Need a recommendation? Use Buy Checking a target model? Use Fit Want a setup answer? Use Run Browse Macs by exact hardware Need the full audit trail? Use Bench

Current published coverage

Published model coverage includes Devstral Small 1.1, Qwen3.5-397B-A17B, Qwen 3 235B-A22B, GLM-5 plus 2 more published models. Published runtime coverage on this chip includes LM Studio, MLX. Fastest published row is 43.0 tok/s on Devstral Small 1.1 at 4bit. Largest published footprint on this page is 415.4 GB for GLM-5. Longest published context is 128k.

Devstral Small 1.1 Qwen3.5-397B-A17B Qwen 3 235B-A22B GLM-5 Llama 3.3 70B Gemma 3 27B

Other published M3 Ultra variants: 256 GB

Macs shipping with the M3 Ultra (512 GB)

Mac Studio M3 Ultra 96GB Mac Studio M3 Ultra 256GB

This chip is part of a family. Compare all M3 Ultra RAM variants →

Raw benchmark rows for M3 Ultra (512 GB)

Rows sorted by avg tok/s descending. Click source badge to see original measurement page.

Model	Quant	RAM req.	Context	Avg tok/s	Prompt tok/s	Runtime	Source
Devstral Small 1.1	4bit	—	—	43.0 tok/s	—	LM Studio	ref
Qwen3.5-397B-A17B	q4.1bit	—	—	40.2 tok/s	—	MLX	ref
Qwen 3 235B-A22B	4bit	—	—	27.4 tok/s	—	LM Studio	ref
GLM-5	4bit	391.8 GB	1k	16.7 tok/s	187.0 tok/s	MLX	ref
Llama 3.3 70B	4bit	—	8k	15.5 tok/s	150.0 tok/s	LM Studio	ref
GLM-5	4bit	394.1 GB	4k	13.7 tok/s	180.1 tok/s	MLX	ref
GLM-5	4bit	396.7 GB	8k	13.2 tok/s	154.1 tok/s	MLX	ref
GLM-5	4bit	402.7 GB	16k	12.0 tok/s	117.4 tok/s	MLX	ref
Gemma 3 27B	bf16	52.6 GB	128k	11.2 tok/s	—	LM Studio	ref
GLM-5	4bit	415.4 GB	33k	10.7 tok/s	77.7 tok/s	MLX	ref
Llama 3.3 70B	4bit	—	40k	9.6 tok/s	103.0 tok/s	LM Studio	ref
Llama 3.3 70B	8bit	—	8k	8.5 tok/s	150.0 tok/s	LM Studio	ref
Llama 3.3 70B	8bit	—	40k	6.5 tok/s	101.0 tok/s	LM Studio	ref

Models tested on this chip

Devstral Small 1.1 Qwen3.5-397B-A17B Qwen 3 235B-A22B GLM-5 Llama 3.3 70B Gemma 3 27B

Next step

Use Rankings when you need the best overall answer for a Mac, and keep this page open when you need the evidence behind that recommendation.

If you are comparing local Apple Silicon against rented cloud hardware, use AI Data Center Index for current GPU rental context.

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 →