Every memory option compared, 100 GB/s memory bandwidth. Updated 2 November 2026.
M2 / M3 are sold with 8, 16, 24 GB of memory. Counting the 31 models tracked here: 8 GB runs 9 models well; 16 GB runs 15 models well; 24 GB runs 19 models well. Memory bandwidth is 100 GB/s, only about 83% of the memory speed of a base M4 (120 GB/s), so expect the same model to be slower. 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 | 8 GB | 16 GB | 24 GB |
|---|---|---|---|
| LTX-2.3 | No | No | Slow about 14 min |
| Gemma 4 31B | No | No | Slow 2/s |
| DeepSeek R1 32B | No | No | Slow 2/s |
| Qwen3 Coder 30B | No | No | Tight 13/s |
| Qwen 3.6 27B | No | No | Slow 2/s |
| Gemma 4 26B A4B | No | No | Tight 10/s |
| Z-Image Turbo 6B | No | Slow about 3 min | Good about 2 min |
| Wan 2.2 5B | No | Slow about 50 min | Good about 25 min |
| gpt-oss 20B | No | Tight 11/s | Great 12/s |
| ACE-Step 1.5 | No | Good about 5 min | Good about 5 min |
| FLUX.2 klein 4B | No | Good about 1 min | Good about 1 min |
| Phi-4 14B | No | Slow 6/s | Slow 6/s |
| Gemma 4 12B | No | Good 7/s | Good 7/s |
| Stable Diffusion XL | Slow about 1 min | Good about 30 s | Good about 30 s |
| Qwen3 TTS 1.7B | Slow about 3 min | Good about 2 min | Good about 2 min |
| Gemma 4 E4B | Tight 15/s | Great 17/s | Great 17/s |
| Qwen 3.5 9B | Tight 8/s | Good 8/s | Good 8/s |
| Qwen3 VL 8B | Tight 8/s | Good 8/s | Good 8/s |
| DeepSeek R1 8B | Tight 9/s | Good 10/s | Good 10/s |
| Qwen2.5 Coder 7B | Tight 10/s | Good 11/s | Good 11/s |
| Gemma 4 E2B | Great 26/s | Great 26/s | Great 26/s |
| Whisper large-v3 turbo | Good about 2 min | Good about 2 min | Good about 2 min |
| Qwen 3.5 4B | Great 16/s | Great 16/s | Great 16/s |
| Kokoro TTS | Good about 20 s | Good about 20 s | Good about 20 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 8 GB is usable by the GPU. The largest chat or coding model that fits is Gemma 4 E4B (7.4 GB), generating about 15 words a second with 0.2 GB to spare. The quickest is Gemma 4 E2B at about 26 words a second. Closest miss: Gemma 4 12B, which needs about 8.9 GB (1 GB more than this configuration can give the GPU).
Find M3 Macs with 8 GB on Amazon
About 15 GB is usable by the GPU. The largest chat or coding model that fits is gpt-oss 20B (15.2 GB), generating about 11 words a second with no spare room. The quickest is Gemma 4 E2B at about 26 words a second. Closest miss: Gemma 4 26B A4B, which needs about 19.4 GB (4 GB more than this configuration can give the GPU).
Find M3 Macs with 16 GB on Amazon
About 23 GB is usable by the GPU. The largest chat or coding model that fits is Qwen3 Coder 30B (20.4 GB), generating about 13 words a second with 2.4 GB to spare. The quickest is Gemma 4 E2B at about 26 words a second. Closest miss: Qwen 3.5 35B A3B, which needs about 23.6 GB (1 GB more than this configuration can give the GPU).
For everyday chat, 8 GB runs 9 of 31 models well. 16 GB to 24 GB unlocks 4 more: Gemma 4 26B A4B, Qwen3 Coder 30B, Z-Image Turbo 6B, Wan 2.2 5B.
No. It needs about 21.5 GB of usable memory and this chip's largest option is 24 GB.
8 GB cannot run gpt-oss 20B; 16 GB runs gpt-oss 20B at about 11 words a second; 24 GB runs gpt-oss 20B at about 12 words a second.
As an Amazon Associate I earn from qualifying purchases.
Unlocks 8 more: Qwen 3.5 35B A3B, Gemma 4 31B, Qwen 3.6 27B, Qwen 3.6 35B A3B and more.
Unlocks 11 more: Qwen 3.5 35B A3B, Gemma 4 31B, Llama 3.3 70B, Qwen 3.5 122B A10B and more.