Only with CPU offload
Llama 4 Scout 109B-A17B at Q4_K_M needs 63.4 GB but only 62.4 GB is addressable, so about 2% of the layers would stream from system RAM at roughly 60 GB/s. Expect around 28 tokens per second — usable for batch work, painful for chat.
A 10M-token window on paper, chunked attention in practice (8K chunks on 3 of 4 layers). Needs 64 GB+ at Q4.
What hardware do I need for Llama 4 Scout 109B-A17B? →
Fits instead: Llama 3.3 70B Instruct (42.8 GB) · Llama 3.1 8B Instruct (6.1 GB)
The VRAM budget
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| Q8_0 | 107.9 GB | 110.0 GB | — | ~1.8 | −0.1% ppl | 47.6 GB over |
| Q6_K | 83.2 GB | 85.3 GB | — | ~3.6 | −0.4% ppl | 22.9 GB over |
| Q5_K_M | 71.9 GB | 74.0 GB | — | ~6.5 | −0.8% ppl | 11.6 GB over |
| Q4_K_M | 61.3 GB | 63.4 GB | 2K | ~28 | −1.9% ppl | 1.0 GB over |
| Q3_K_M | 49.6 GB | 51.7 GB | 235K | 49 | −5.4% ppl | Long context |
| Q2_K | 42.5 GB | 44.6 GB | 387K | 58 | −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
$ llama-server \
-hf meta-llama/Llama-4-Scout-17B-16E-Instruct:Q4_K_M \
-c 8192 -ngl 47
The engine underneath most of the others. Every knob is exposed. More on llama.cpp.