Is the M3 Ultra faster than the M2 Ultra for LLMs?

About the same: their memory bandwidth is within a few percent. The M3 Ultra goes up to 512 GB of memory, the M2 Ultra to 192 GB.

Which can run bigger models, the M2 Ultra or the M3 Ultra?

The M3 Ultra: The M3 Ultra goes up to 512 GB of memory, the M2 Ultra to 192 GB. At its largest it runs 60 of 73 open models well, against 49.

M2 Ultra vs M3 Ultra for local LLMs: memory, speed, which to buy

Side by side

	M2 Ultra	M3 Ultra
Memory sizes	64, 128, 192 GB	96, 256, 512 GB
Memory bandwidth	800 GB/s	819 GB/s
For models, at most	about 144 GB	about 384 GB
Models that run well	49 of 73	60 of 73
MLX seat class	Ultra	Ultra
Found in	Mac Studio (2023), Mac Pro (2023), Mac Pro (Rack, 2023)	Mac Studio (2025)

Speed with popular models

Tokens a second for one conversation, each chip at its largest memory size (192 GB and 512 GB), with the best version of the model that fits.

Model	M2 Ultra	M3 Ultra
Llama 3.1 8B Instruct8B	42–75 8-bit	43–77 8-bit
gpt-oss-20b21B · 3.6B active	189–337 native 4.25-bit	194–345 native 4.25-bit
Qwen3.6 35B A3B35B · 3B active	113–200 8-bit	115–205 8-bit
Qwen3.8 27B27B	13–22 8-bit	13–23 8-bit
Llama 3.3 70B Instruct70B	5–9 8-bit	5–9 8-bit
gpt-oss-120b117B · 5.1B active	133–237 native 4.25-bit	137–243 native 4.25-bit
Qwen3.5 122B A10B122B · 10B active	34–60 8-bit	35–62 8-bit
DeepSeek V4 Flash 0731284B · 13B active	63–112 3-bit lower quality	27–47 8-bit

Questions

Is the M3 Ultra faster than the M2 Ultra for LLMs?: About the same: their memory bandwidth is within a few percent. The M3 Ultra goes up to 512 GB of memory, the M2 Ultra to 192 GB.
Which can run bigger models, the M2 Ultra or the M3 Ultra?: The M3 Ultra: The M3 Ultra goes up to 512 GB of memory, the M2 Ultra to 192 GB. At its largest it runs 60 of 73 open models well, against 49.

More comparisons

The M2 Ultra for local AI · The M3 Ultra for local AI · Compare every model · The best Mac for local LLMs

How it’s worked out

Memory: the model’s parameters at 8, 6, 4, 3 or 2 bits per weight plus MLX’s group scales (half a bit per weight), or its native format, with room for the context on top. By default macOS lets the GPU use about 75% of unified memory. Speed: memory bandwidth divided by the bytes read per token is the ceiling for one conversation; in llama.cpp’s public Apple Silicon benchmarks real runs reach about 45% to 80% of it, which is the range shown. These are estimates, not measurements: measure your own Mac on MLX Bench. Model figures from each official model card, chip figures from Apple’s Tech Specs. Checked 2026-10-19. Same data as mlx.app/compare.