Only with CPU offload
Mistral Small 4 119B-A6B at Q4_K_M needs 67.7 GB but only 22.4 GB is addressable, so about 68% of the layers would stream from system RAM at roughly 60 GB/s. Expect around 5.5 tokens per second — usable for batch work, painful for chat.
Instruct, reasoning, vision and code in one 119B MoE. A latent KV cache (320 wide) keeps context cheap; the weights still want 64 GB+.
What hardware do I need for Mistral Small 4 119B-A6B? →
Fits instead: Mistral Small 3.2 24B (15.1 GB)
The VRAM budget
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| Q8_0 | 117.8 GB | 118.5 GB | — | ~2.6 | −0.1% ppl | 96.1 GB over |
| Q6_K | 90.9 GB | 91.7 GB | — | ~3.6 | −0.4% ppl | 69.3 GB over |
| Q5_K_M | 78.5 GB | 79.3 GB | — | ~4.4 | −0.8% ppl | 56.9 GB over |
| Q4_K_M | 66.9 GB | 67.7 GB | — | ~5.5 | −1.9% ppl | 45.3 GB over |
| Q3_K_M | 54.2 GB | 54.9 GB | — | ~7.6 | −5.4% ppl | 32.5 GB over |
| Q2_K | 46.4 GB | 47.2 GB | — | ~9.8 | −15% ppl | 24.8 GB over |
Quality is the published perplexity delta against f16 weights. Max context assumes an f16 KV cache; q8_0 roughly doubles it. This model uses multi-head latent attention, so its cache is a compressed latent rather than full K and V.
How to run it
$ llama-server \
-hf mistralai/Mistral-Small-4-119B-2603:Q4_K_M \
-c 8192 -ngl 11
The engine underneath most of the others. Every knob is exposed. More on llama.cpp.