Side by side
Speed with popular models
Tokens a second for one conversation, each chip at its largest memory size (128 GB and 192 GB), with the best version of the model that fits.
| Model | M1 Ultra | M2 Ultra |
|---|---|---|
| Llama 3.1 8B Instruct8B | 42–75 8-bit | 42–75 8-bit |
| gpt-oss-20b21B · 3.6B active | 189–337 native 4.25-bit | 189–337 native 4.25-bit |
| Qwen3.6 35B A3B35B · 3B active | 113–200 8-bit | 113–200 8-bit |
| Qwen3.8 27B27B | 13–22 8-bit | 13–22 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 | 133–237 native 4.25-bit |
| Qwen3.5 122B A10B122B · 10B active | 64–114 4-bit | 34–60 8-bit |
| DeepSeek V4 Flash 0731284B · 13B active | 88–156 2-bit lower quality | 63–112 3-bit lower quality |
Questions
- Is the M2 Ultra faster than the M1 Ultra for LLMs?
- About the same: their memory bandwidth is within a few percent. The M2 Ultra goes up to 192 GB of memory, the M1 Ultra to 128 GB.
- Which can run bigger models, the M1 Ultra or the M2 Ultra?
- The M2 Ultra: The M2 Ultra goes up to 192 GB of memory, the M1 Ultra to 128 GB. At its largest it runs 49 of 73 open models well, against 47.
More comparisons
The M1 Ultra for local AI · The M2 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.