Only with CPU offload
Devstral Small 2 24B at Q4_K_M needs 15.3 GB but only 14.4 GB is addressable, so about 7% of the layers would stream from system RAM at roughly 60 GB/s. Expect around 18 tokens per second — usable for batch work, painful for chat.
93% on GPU
8K context
Q4_K_M · 13.5 GB
Apache 2.0
Released Dec 2025
Vision
Not in the Ollama library
Built for software-engineering agents (OpenHands, Cline). Dense 24B, 384K window. Not in the Ollama library.
The VRAM budget
weights 13.5 GB
Weights 13.5 GB
KV cache @ 8K 1.25 GB
Runtime overhead 0.6 GB
Over budget 0.9 GB past 14.4 GB
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| Q8_0 | 23.7 GB | 25.6 GB | — | ~3.0 | −0.1% ppl | 11.2 GB over |
| Q6_K | 18.3 GB | 20.2 GB | — | ~5.3 | −0.4% ppl | 5.8 GB over |
| Q5_K_M | 15.8 GB | 17.7 GB | — | ~8.4 | −0.8% ppl | 3.3 GB over |
| Q4_K_M | 13.5 GB | 15.3 GB | 1K | ~18 | −1.9% ppl | 0.9 GB over |
| Q3_K_M | 10.9 GB | 12.8 GB | 18K | 40 | −5.4% ppl | Long context |
| Q2_K | 9.4 GB | 11.2 GB | 28K | 46 | −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
$ llama-server \
-hf mistralai/Devstral-Small-2-24B-Instruct-2512:Q4_K_M \
-c 8192 -ngl 37
The engine underneath most of the others. Every knob is exposed. More on llama.cpp.
01Download is 13.5 GB. Keep it on an SSD — a first load off a spinning disk takes minutes.
02Close anything else holding VRAM. A browser with hardware acceleration can sit on 1–2 GB.
03If it falls back to CPU silently, drop the context first, then step down a quantisation.