Can I run Codestral 22B?

Codestral 22B is Mistral AI's dedicated coding model, built for code completion and generation rather than open-ended chat. At 22.2 billion parameters it is a real step up in size from the 7B-12B models most people start with, and the footprint shows it: a 4-bit quant lands around 13.5 GB and the model wants at least 24 GB of RAM to run comfortably. That puts it out of reach of a 12 GB card like an RTX 3060, where it simply does not fit. A 24 GB GPU such as an RTX 4090, or an Apple Silicon Mac with plenty of unified memory, is the realistic entry point.

On an RTX 4090 you can expect around 64 tokens per second at 4-bit, fast enough that code suggestions feel close to instant. An Apple M-series Max sits lower at roughly 26 tok/s, still usable for interactive editing, while CPU-only on DDR5 drops to about 4 tok/s, which is really only for batch jobs you can walk away from. The context window is 32K, modest by today's standards but generous for most single-file and small-repo work. Fill it completely and you are looking at about 20.1 GB of total memory, so on a 24 GB card a full context leaves little headroom.

Against its stablemates, Codestral is the specialist: Mistral Nemo 12B is the lighter, general-purpose chat pick if you are short on memory, and GPT-OSS 20B generally has the edge for reasoning and conversation. Codestral's standout trait is that it was trained specifically for code across many languages, so for fill-in-the-middle completion and generation it tends to feel more on-target than a same-size generalist. The serious caveat is the license: Codestral ships under Mistral's MNPL, a non-production license, so you cannot use it commercially or in a production product. Treat it as a research and personal-use tool only.