Codestral on NVIDIA GeForce RTX 3060 · 12GB VRAM

⚠️ Partial GPU offload — Codestral 22B @ Q4_K_M
Codestral 22B at Q4_K_M needs ~14.9 GB (weights 12.6 GB + KV 1.8 GB + overhead 643 MB @ 8K ctx) of your 10.8 GB usable VRAM — some layers spill to system RAM. Expect ~3–5 tok/s (slow).
context @ Q4_K_M: runs to its full 32K
quantneedsspeed
~ Q8_024.9 GB—
~ Q5_K_M17.2 GB—
~ Q4_K_M14.9 GB~3–5 tok/s (slow)
~ Q3_K_M12.4 GB~4–6 tok/s (slow)
$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: Partial GPU offload
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 3060: Qwen3 · DeepSeek-R1 (distill) · Qwen2.5 · Llama 3.1 · Gemma 3

Best models for a NVIDIA GeForce RTX 3060 · 12GB 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 3060 · 12GB VRAM too:

modelsharesverdict here
Qwen2.5-Coder 8kv/128hd ✅ Runs (tight)
Llama 3.2 8kv/128hd ✅ Runs comfortably
Llama 3.3 8kv/128hd ❌ Won't fit
Mistral 7B 8kv/128hd ✅ Runs (tight)
Mistral Small 3 8kv/128hd ⚠️ Partial GPU offload
Mistral Nemo 8kv/128hd ✅ Runs (tight)
Llama 3.2 Vision 8kv/128hd ✅ Runs comfortably
Mixtral 8x7B 8kv/128hd 🐢 CPU-only (slow)

…and 7 more — see the Codestral page.

Check any combo yourself: open the checker →