M5 Ultra vs M3 Ultra: local AI, measured

Reading got about 4x faster. Writing got about 1.5x.

A chat model does two different jobs. First it reads your prompt, all of it at once, as huge matrix multiplications. Then it writes the answer one token at a time, and for every token a dense model reads essentially every one of its weights out of memory again, plus the cached keys and values for the conversation so far.

Reading is limited by math. The M5 Ultra puts Neural Accelerators in every GPU core, which Apple says "deliver dramatically faster matrix multiplication" (Apple Newsroom). On the dense models, reading came out 3.9x to 4.1x faster.

Writing is limited by memory bandwidth. Apple says that went from 819 GB/s to 1.2 TB/s, which is 1.47x. Writing came out 1.5x faster on four of the five models, and 1.8x on Qwen3 14B. That is an explanation of the numbers, not a measurement of the chip: the full reasoning.

How many times faster the M5 Ultra was at each half of the job, 32K-token prompt, median of 5 runs. The vertical line is Apple's memory bandwidth ratio. data/language-models/summary.json

Dense models gained more than mixture-of-experts models

A dense model runs every token through every weight. A mixture-of-experts model sends each token to a few small "experts" and leaves the rest idle, so a 35B model like Qwen3.6 35B-A3B only uses about 3B per token.

On reading, the dense models gained 3.9x to 4.1x and the mixture-of-experts models 2.7x to 3.0x. The likely explanation (not profiled): a mixture-of-experts model spends part of its reading time routing tokens between experts and on attention, which is not the big matrix math the new hardware speeds up most. The 35B-A3B mixture-of-experts model is still the fastest model here on both Macs. Dense vs mixture of experts, explained.

Model	Reading, tok/s		Writing, tok/s		First word
Qwen3 14B14B, all active per token · 4-bit	551 → 2,278	4.1x	31.5 → 55.4	1.8x	59.0 s → 14.3 s	4.1x
Qwen3.8 27B27B, all active per token · 4-bit	388 → 1,526	3.9x	27.8 → 42.6	1.5x	83.6 s → 21.3 s	3.9x
Gemma 4 31B31B, all active per token · 16-bit	294 → 1,217	4.1x	9.0 → 13.6	1.5x	110.4 s → 26.7 s	4.1x
Qwen3.6 35B-A3B35B total, 3B active per token · 4-bit	2,193 → 6,012	2.7x	71.5 → 109.1	1.5x	14.8 s → 5.40 s	2.7x
Qwen3.5 122B-A10B122B total, 10B active per token · 4-bit	771 → 2,300	3.0x	42.3 → 65.3	1.5x	42.2 s → 14.1 s	3.0x

The wait for the first word

With a long prompt, most of what you feel is the wait before the answer starts. That is the reading half, so it shrank the most.

Seconds from sending a 32K-token prompt to the first word, median of 5 runs. Shorter is better. data/language-models/summary.json

The longer the prompt, the more of the new chip you get

A short prompt is over almost before it starts, and much of its time is fixed overhead. A long one is mostly matrix math, so the gap opens up. Pasting in a document, a codebase or a long chat is where the M5 Ultra pulls away.

Reading speed of Qwen3.8 27B 4-bit, prompt tokens per second, median of 5 runs. data/language-models/summary.json

Reading gain	512	4K	16K	32K
Qwen3 14Bdense	2.8x	4.1x	4.2x	4.1x
Qwen3.8 27Bdense	2.9x	3.9x	4.0x	3.9x
Gemma 4 31Bdense	2.9x	4.0x	4.2x	4.1x
Qwen3.6 35B-A3Bmixture of experts	1.6x	2.5x	2.7x	2.7x
Qwen3.5 122B-A10Bmixture of experts	2.0x	2.9x	3.0x	3.0x

Everyday AI work

Image generation is almost pure matrix math and gained the most. Transcription reads and then writes, so it lands in between. Hardware video transcoding runs on the media engines, not the GPU, and gained the least. A tiny decision model answering one short question barely changed.

M3 Ultra time divided by M5 Ultra time. Images are the median of 5 runs; the rest are single runs. data/everyday.json, data/images/summary.json, data/transcription.json

Every result

Test	M3 Ultra	M5 Ultra	M5 Ultra is
Live race, Qwen3.8 27B, 32K prompt: wait for the first word	73.8 s	18.5 s	4.0x	details
Live race, same race: writing speed	28.3 tok/s	43.6 tok/s	1.5x	details
Live race, same race: whole answer	80.4 s	22.8 s	3.5x	details
Live race, a short question: whole answer	3.42 s	2.10 s	1.6x	details
One 1024x1024 image, FLUX.2 klein 4B (median of 5)	7.00 s	1.77 s	4.0x	details
One 1024x1024 image, Z-Image Turbo (median of 5)	27.9 s	9.25 s	3.0x	details
4-second video clip with sound, LTX-2.5 (load included)	48.4 s	17.8 s	2.7x	details
2-second video clip with sound, LTX-2.5 (load included)	24.3 s	14.0 s	1.7x	details
Describe 60 s of footage, Qwen3-VL 32B (load included)	61.4 s	23.5 s	2.6x	details
Describe 60 s of footage, Qwen3-VL 8B (load included)	17.2 s	7.0 s	2.5x	details
Transcode a 13.8-min 4K take to ProRes 422 (hardware)	91 s	47 s	1.9x	details
Transcode the same take to HEVC (hardware)	120 s	90 s	1.3x	details
Laya, 300 decision questions (load included)	35.6 s	24.2 s	1.5x	details
Kev-4B, one question, model warm	~0.20 s	~0.19 s	about 1.0x	details
Kev-4B, the first (cold) question	3.88 s	3.98 s	0.97x	details
Hermes agent on the 27B: read a repo, answer five questions	not run	89 s	—	details
Resolve edit A: the 27B decides 30 keep ranges and builds the timeline	not run	about 2 min	—	details
Resolve edit B: Kev judges 143 segments, the 27B builds	not run	95 s	—	details

Transcription, all six Whisper models

Whisper model	13.8-min take		2 h 5 min shoot day
Whisper large-v3	19.63 s → 11.21 s	1.75x	162.2 s → 90.0 s	1.80x
Whisper large-v3 turbo	7.21 s → 3.05 s	2.36x	57.8 s → 27.3 s	2.12x
Whisper medium	12.44 s → 7.44 s	1.67x	98.7 s → 59.8 s	1.65x
Whisper small	5.88 s → 3.79 s	1.55x	46.8 s → 30.9 s	1.52x
Whisper base	3.00 s → 2.06 s	1.46x	25.6 s → 17.6 s	1.46x
Whisper tiny	2.45 s → 1.66 s	1.48x	18.5 s → 12.7 s	1.45x

What the Macs made

Pictures made by two AI image models on each Mac, from the same prompt and seed. They are not bit-identical, and each time is the median per image with the model already loaded.

FLUX.2 klein 4B on the M3 Ultra: A cutaway illustration of a desktop computer on a workbench, warm afternoon light, fine mechanical detail, shallow depth of field — **M3 Ultra** FLUX.2 klein 4B, 7.00 s per image

FLUX.2 klein 4B on the M5 Ultra: A cutaway illustration of a desktop computer on a workbench, warm afternoon light, fine mechanical detail, shallow depth of field — **M5 Ultra** FLUX.2 klein 4B, 1.77 s per image

Z-Image Turbo on the M3 Ultra: A cutaway illustration of a desktop computer on a workbench, warm afternoon light, fine mechanical detail, shallow depth of field — **M3 Ultra** Z-Image Turbo, 27.9 s per image

Z-Image Turbo on the M5 Ultra: A cutaway illustration of a desktop computer on a workbench, warm afternoon light, fine mechanical detail, shallow depth of field — **M5 Ultra** Z-Image Turbo, 9.25 s per image

Machine generated. These clips were made by the LTX-2.5 video model through an experimental MLX port. They are not real footage (LTX-2.x Community License). Times include loading the model.

M3 Ultra 48.4 s, 4-second clip

M5 Ultra 17.8 s, 4-second clip

M3 Ultra 24.3 s, 2-second clip

M5 Ultra 14.0 s, 2-second clip

The race page during a 32K-prompt race: the M5 Ultra is already writing while the M3 Ultra is still reading the prompt — An earlier 32K-prompt race, caught mid-race: the same request went to both Macs at the same instant, and the M5 Ultra is already writing while the M3 Ultra is still reading. This is a different run from the live race in the results table. About the race

The two machines

The new Mac has the faster chip. The old Mac has twice the memory, which decides how big a model fits. Every model we tested fit on both. The two Macs side by side.

	M3 Ultra 512GB	M5 Ultra 256GB
Chip	M3 Ultra	M5 Ultra
CPU cores	32 (24 performance + 8 efficiency)	36 (12 super + 24 performance)
GPU cores	80	80
Memory in this unit	512 GB	256 GB
Memory bandwidth, Apple says	819 GB/s	1,200 GB/s
macOS as tested	26.5.1	27.0
Whose	Chuck's own	loaned by Apple for review

How we measured

Chat models ran on mlx_lm.server (mlx-lm 0.31.3, MLX 0.32.2) on each Mac, prompt cache off, thinking off, temperature 0.
Every cell: 1 warm-up run thrown away, then 5 measured runs, 30 s of rest between cells, 128-token answers. Headline numbers are medians, never means.
Reading speed is true prompt tokens divided by the seconds to the first word. Writing speed is (output tokens − 1) divided by the time from the first piece of the answer to the last.
Every prompt starts with a unique run id and every run logged the server's cached tokens (all zero), so no answer came from a cache.
Images: mflux 0.19.2 on MLX 0.32.2, 1024x1024, the same prompt and seed on both Macs, 5 measured images after a warm-up, with the model load timed separately.
Everything else ran once per Mac, usually on both Macs at the same moment, the way the video shows it.
The Macs ran different macOS versions (26.5.1 and 27.0), and a stuck background process held one M3 Ultra CPU core during most tests. It can only have made the M3 Ultra look slower. See docs/limits.md.

More: the method, honest limits, Apple's claims next to ours, glossary.

Get the data, run it yourself

Every number on this page is in data/ as JSON, with the raw logs beside it (schema). The benchmark tool is in harness/, and it runs with no AI hardware at all against two fake Macs:

git clone https://github.com/theNetworkChuck/mac-studio-m5-ultra-local-ai.git
cd mac-studio-m5-ultra-local-ai/harness
bun run mock    # terminal 1: two fake Macs
bun run race    # terminal 2: open the URL it prints, press START

To test your own Macs: docs/reproduce.md. This page and the README tables are generated from the data by scripts/build.py.