Yes — with 11.3 GB to spare
Ministral 3 8B at Q4_K_M fits your M3 · 24 GB entirely in unified memory at 8K context, at an estimated 11 tokens per second. Past 93K the KV cache pushes it over — quantise the cache to q8_0, or step down a quantisation, to go longer.
Fully in unified memory
8K context
Q4_K_M · 5.0 GB
Apache 2.0
Released Dec 2025
Vision
Mistral's 8B with images in. Plain GQA, so budget more KV cache than Qwen3.5 9B at the same context.
The VRAM budget
weights 5.0 GB
Weights 5.0 GB
KV cache @ 8K 1.06 GB
Runtime overhead 0.6 GB
Free 11.3 GB of 18.0 GB
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| F16 | 16.6 GB | 18.3 GB | 5K | ~3.4 | Reference | 0.3 GB over |
| Q8_0 | 8.8 GB | 10.5 GB | 64K | 6.3 | −0.1% ppl | Long context |
| Q6_K | 6.8 GB | 8.5 GB | 79K | 8.2 | −0.4% ppl | Long context |
| Q5_K_M | 5.9 GB | 7.6 GB | 86K | 9.5 | −0.8% ppl | Long context |
| Q4_K_M | 5.0 GB | 6.7 GB | 93K | 11 | −1.9% ppl | Recommended |
| Q3_K_M | 4.1 GB | 5.7 GB | 100K | 14 | −5.4% ppl | Long context |
| Q2_K | 3.5 GB | 5.1 GB | 104K | 16 | −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/Ministral-3-8B-Instruct-2512-4bit \ --max-tokens 512 --prompt "Hello"
Apple's own array framework. The fastest path on Apple Silicon. More on MLX.
01Download is 5.0 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 18.0 GB of 24 GB. Raising it with iogpu.wired_limit_mb is possible, and risky.
03There is room to go to 93K context on this card.