Yes — with 90.3 GB to spare
Mistral 7B Instruct v0.3 at Q4_K_M fits your M3 Max · 128 GB entirely in unified memory at 8K context, at an estimated 55 tokens per second. There is room for its full 32K window.
Fully in unified memory
8K context
Q4_K_M · 4.1 GB
Apache 2.0
Released May 2024
Old but extremely well behaved, and permissively licensed for commercial use.
The VRAM budget
weights 4.1 GB
Weights 4.1 GB
KV cache @ 8K 1.00 GB
Runtime overhead 0.6 GB
Free 90.3 GB of 96.0 GB
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| F16 | 13.5 GB | 15.1 GB | 32K | 17 | Reference | Long context |
| Q8_0 | 7.2 GB | 8.8 GB | 32K | 31 | −0.1% ppl | Long context |
| Q6_K | 5.5 GB | 7.1 GB | 32K | 40 | −0.4% ppl | Long context |
| Q5_K_M | 4.8 GB | 6.4 GB | 32K | 47 | −0.8% ppl | Long context |
| Q4_K_M | 4.1 GB | 5.7 GB | 32K | 55 | −1.9% ppl | Recommended |
| Q3_K_M | 3.3 GB | 4.9 GB | 32K | 68 | −5.4% ppl | Long context |
| Q2_K | 2.8 GB | 4.4 GB | 32K | 79 | −15% ppl | Long context |
Quality is the published perplexity delta against f16 weights. Max context assumes an f16 KV cache; q8_0 roughly doubles it.
How to run it
$ pip install mlx-lm $ mlx_lm.generate --model mlx-community/Mistral-7B-Instruct-v0.3-4bit \ --max-tokens 512 --prompt "Hello"
Apple's own array framework. The fastest path on Apple Silicon. More on MLX.
01Download is 4.1 GB. Keep it on an SSD — a first load off a spinning disk takes minutes.
02macOS caps what the GPU may wire down at about 96.0 GB of 128 GB. Raising it with iogpu.wired_limit_mb is possible, and risky.
03There is room to go to the model's full 32K context on this card.