Every memory option compared, 819 GB/s memory bandwidth. Updated 2 November 2026.
M3 Ultra is sold with 96, 256, 512 GB of memory. Counting the 31 models tracked here: 96 GB runs 30 models well; 256 GB runs 30 models well; 512 GB runs 31 models well. Memory bandwidth is 819 GB/s, about 6.8 times the memory speed of a base M4 (120 GB/s), so the same model generates words proportionally faster. Apple Silicon shares one pool of memory between the processor and the GPU, and macOS lets the GPU use about two thirds of it on Macs up to 36 GB and about three quarters above that, so the memory you buy decides which models load at all.
| Model | 96 GB | 256 GB | 512 GB |
|---|---|---|---|
| DeepSeek V3 671B | No | No | Tight 11/s |
| Qwen 3.5 122B A10B | Tight 38/s | Great 41/s | Great 41/s |
| gpt-oss 120B | Great 79/s | Great 79/s | Great 79/s |
| Llama 3.3 70B | Good 10/s | Good 10/s | Good 10/s |
| Wan 2.2 14B | Great about 8 min | Great about 8 min | Great about 8 min |
| LTX-2.3 | Great about 1 min | Great about 1 min | Great about 1 min |
| Qwen 3.6 35B A3B | Great 123/s | Great 123/s | Great 123/s |
| Qwen 3.5 35B A3B | Great 128/s | Great 128/s | Great 128/s |
| Gemma 4 31B | Great 22/s | Great 22/s | Great 22/s |
| DeepSeek R1 32B | Great 22/s | Great 22/s | Great 22/s |
| Qwen3 Coder 30B | Great 118/s | Great 118/s | Great 118/s |
| Qwen 3.6 27B | Great 23/s | Great 23/s | Great 23/s |
| Gemma 4 26B A4B | Great 92/s | Great 92/s | Great 92/s |
| Z-Image Turbo 6B | Great about 15 s | Great about 15 s | Great about 15 s |
| Wan 2.2 5B | Great about 4 min | Great about 4 min | Great about 4 min |
| gpt-oss 20B | Great 98/s | Great 98/s | Great 98/s |
| ACE-Step 1.5 | Great about 47 s | Great about 47 s | Great about 47 s |
| FLUX.2 klein 4B | Great about 11 s | Great about 11 s | Great about 11 s |
| Phi-4 14B | Great 47/s | Great 47/s | Great 47/s |
| Gemma 4 12B | Great 54/s | Great 54/s | Great 54/s |
| Stable Diffusion XL | Great about 7 s | Great about 7 s | Great about 7 s |
| Qwen3 TTS 1.7B | Great about 17 s | Great about 17 s | Great about 17 s |
| Gemma 4 E4B | Great 134/s | Great 134/s | Great 134/s |
| Qwen 3.5 9B | Great 65/s | Great 65/s | Great 65/s |
| Qwen3 VL 8B | Great 71/s | Great 71/s | Great 71/s |
| DeepSeek R1 8B | Great 83/s | Great 83/s | Great 83/s |
| Qwen2.5 Coder 7B | Great 92/s | Great 92/s | Great 92/s |
| Gemma 4 E2B | Great 190+/s | Great 190+/s | Great 190+/s |
| Whisper large-v3 turbo | Great about 25 s | Great about 25 s | Great about 25 s |
| Qwen 3.5 4B | Great 131/s | Great 131/s | Great 131/s |
| Kokoro TTS | Great about 4 s | Great about 4 s | Great about 4 s |
Text models show words a second. Image, video and voice models show the time per job (image, 4 second clip, minute of speech or similar).
As an Amazon Associate I earn from qualifying purchases.
About 91 GB is usable by the GPU. The largest chat or coding model that fits is Qwen 3.5 122B A10B (85.5 GB), generating about 38 words a second with 5.7 GB to spare. The quickest is Gemma 4 E2B at about 190+ words a second. Closest miss: DeepSeek V3 671B, which needs about 424.7 GB (334 GB more than this configuration can give the GPU).
Find M3 Ultra Macs with 96 GB on Amazon
About 243 GB is usable by the GPU. The largest chat or coding model that fits is Qwen 3.5 122B A10B (85.5 GB), generating about 41 words a second with 157.7 GB to spare. The quickest is Gemma 4 E2B at about 190+ words a second. Closest miss: DeepSeek V3 671B, which needs about 424.7 GB (182 GB more than this configuration can give the GPU).
Find M3 Ultra Macs with 256 GB on Amazon
About 486 GB is usable by the GPU. The largest chat or coding model that fits is DeepSeek V3 671B (424.7 GB), generating about 11 words a second with 61.7 GB to spare. The quickest is Gemma 4 E2B at about 190+ words a second. Nothing popular is out of reach.
For everyday chat, 96 GB runs 30 of 31 models well. 256 GB to 512 GB unlocks 1 more: DeepSeek V3 671B.
Yes, from 96 GB of memory, at about 22 words a second.
96 GB runs gpt-oss 20B at about 98 words a second; 256 GB runs gpt-oss 20B at about 98 words a second; 512 GB runs gpt-oss 20B at about 98 words a second.
You are already at the top end. Nothing popular is out of reach except the giant server models.