Codestral on NVIDIA GeForce RTX 4090 · 24GB VRAM

✅ Runs comfortably — Codestral 22B @ Q5_K_M
Codestral 22B at Q5_K_M needs ~17.2 GB (weights 14.7 GB + KV 1.8 GB + overhead 755 MB @ 8K ctx) of your 21.6 GB usable VRAM — plenty of headroom. Expect ~30–45 tok/s (fast).
context @ Q5_K_M: comfortable to 8K · runs to its full 32K
quantneedsspeed
~ Q8_024.9 GB~6–9 tok/s (slow)
✓ Q5_K_M17.2 GB~30–45 tok/s (fast)
✓ Q4_K_M14.9 GB~30–55 tok/s (fast)
✓ Q3_K_M12.4 GB~40–65 tok/s (very fast)
$runlocal install codestral:22b
⚠ NVIDIA support is best-effort in v0.1 — verify before relying on it.
runlocal verdict card
download card share on x share on reddit drop it in your model-drop post
Codestral: Runs comfortably
for your own README — links back here

Same model, other machines: Apple M1 · Apple M1 Pro · Apple M2 · Apple M3 Pro · Apple M2 Pro · Apple M4 Pro

Also runs on a NVIDIA GeForce RTX 4090: DeepSeek-R1 (distill) · Qwen3 · Mistral Small 3 · Qwen2.5 · Gemma 3

Best models for a NVIDIA GeForce RTX 4090 · 24GB VRAM →

Architecturally similar

These share Codestral's KV-cache geometry — they size memory the same way, so here's how they fit on a NVIDIA GeForce RTX 4090 · 24GB VRAM too:

modelsharesverdict here
Llama 3.1 8kv/128hd ✅ Runs comfortably
Qwen2.5-Coder 8kv/128hd ✅ Runs (tight)
Llama 3.2 8kv/128hd ✅ Runs comfortably
Llama 3.3 8kv/128hd ⚠️ Partial GPU offload
Mistral 7B 8kv/128hd ✅ Runs comfortably
Mistral Nemo 8kv/128hd ✅ Runs comfortably
Llama 3.2 Vision 8kv/128hd ✅ Runs comfortably
Mixtral 8x7B 8kv/128hd ⚠️ Partial GPU offload

…and 8 more — see the Codestral page.

Check any combo yourself: open the checker →