Can I run Mixtral 8x7B (MoE) on Apple M3 Ultra (192GB)?
Yes: Mixtral 8x7B (MoE) runs on Apple M3 Ultra (192GB) with F16 at roughly 20 tokens/s. Estimated memory 89.6 GiB at 8,192 tokens of context.
With F16 you can push the context to about 32,768 tokens before memory runs out (F16 KV cache).
Memory by quantisation (8,192-token context)
| Quantisation | Memory | Verdict | Est. speed |
|---|---|---|---|
| F16Unquantized reference | 89.6 GiB | Fits | ~20 tok/s |
| Q8_0Near-lossless | 48.8 GiB | Fits | ~36 tok/s |
| Q6_KVery close to F16 | 38.3 GiB | Fits | ~46 tok/s |
| Q5_K_MGreat quality/size | 33.5 GiB | Fits | ~52 tok/s |
| Q4_K_MBest default for most | 29.0 GiB | Fits | ~60 tok/s |
| IQ4_XSSmaller 4-bit (imatrix) | 25.7 GiB | Fits | ~67 tok/s |
| Q3_K_MNoticeable loss | 23.9 GiB | Fits | ~72 tok/s |
| Q2_KLast resort; big loss | 20.8 GiB | Fits | ~82 tok/s |
Apple M3 Ultra (192GB) offers about 192 GiB of memory (75% usable for a model) at 819 GB/s. Speed is a bandwidth-bound estimate of single-stream generation at 65% efficiency; real results vary by runtime, drivers and settings, so measure on your machine.
How this is calculated
Weights ≈ parameters × bits per weight ÷ 8. KV cache ≈ 2 × 32 layers × 8 KV heads × 128 head dim × 2 bytes per token. Decode speed ≈ efficiency × memory bandwidth ÷ bytes read per token (this is a mixture-of-experts model, so only the active parameters are read, but all of them must fit in memory). Learn the maths in VRAM math and memory bandwidth, or try the VRAM calculator.
Mixtral 8x7B (MoE) on other hardware
Similar models on Apple M3 Ultra (192GB)
Running slowly or out of memory? See CUDA out of memory and the quantisation lesson.