Yes, just — 8.6 GB spare
Llama 4 Scout 109B-A17B at Q4_K_M fits your M2 Max · 96 GB entirely in unified memory at 8K context, at an estimated 11 tokens per second. Past 191K 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 · 61.3 GB
Llama 4 Community
Released 5 Apr 2025
Vision
A 10M-token window on paper, chunked attention in practice (8K chunks on 3 of 4 layers). Needs 64 GB+ at Q4.
The VRAM budget
weights 61.3 GB
Weights 61.3 GB
KV cache @ 8K 1.50 GB
Runtime overhead 0.6 GB
Free 8.6 GB of 72.0 GB
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| Q8_0 | 107.9 GB | 110.0 GB | — | ~6.2 | −0.1% ppl | 38.0 GB over |
| Q6_K | 83.2 GB | 85.3 GB | — | ~8.0 | −0.4% ppl | 13.3 GB over |
| Q5_K_M | 71.9 GB | 74.0 GB | — | ~9.3 | −0.8% ppl | 2.0 GB over |
| Q4_K_M | 61.3 GB | 63.4 GB | 191K | 11 | −1.9% ppl | Recommended |
| Q3_K_M | 49.6 GB | 51.7 GB | 440K | 13 | −5.4% ppl | Long context |
| Q2_K | 42.5 GB | 44.6 GB | 592K | 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. This model interleaves sliding-window layers (8192 tokens, 12 of 48 layers global), which is why its cache barely grows with context.
How to run it
$ pip install mlx-lm $ mlx_lm.generate --model mlx-community/Llama-4-Scout-17B-16E-Instruct-4bit \ --max-tokens 512 --prompt "Hello"
Apple's own array framework. The fastest path on Apple Silicon. More on MLX.
01Download is 61.3 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 72.0 GB of 96 GB. Raising it with iogpu.wired_limit_mb is possible, and risky.
03Only 8.6 GB is spare, so a long system prompt can still push it over. Its real ceiling here is 191K context.