Codestral 22B: Code Generation Analysis

Codestral 22B hardware requirements

Codestral 22B at Q4_K_M is a 13.3 GB file. It runs on a 16 GB card with a short context and comfortably on a 24 GB card, where there is room for the full 32K window. A 12 GB card can hold the 8.3 GB Q2_K build. Q8_0 is 23.6 GB and needs a 32 GB card. The arithmetic is parameters × bits per weight ÷ 8: 22.25B × 4.8 bits ÷ 8 = 13.3 GB at Q4_K_M.

File sizes are from bartowski/Codestral-22B-v0.1-GGUF and the tag sizes from Ollama's tag list. The context cache is arithmetic from the model's params.json: 56 layers × 8 key-value heads × 128 dimensions, stored twice (keys and values) at 2 bytes, is 229 KB per token — 0.9 GB at 4K context, 1.9 GB at 8K and 7.5 GB at the full 32K. “Memory needed” is file + the 8K cache + a 1 GB allowance for buffers.

With the full 32K context, Q4_K_M needs about 13.3 + 7.5 + 1 = 21.8 GB, which is why 24 GB is the comfortable tier. On a Mac the same totals apply to unified memory, and without a GPU the file loads into system RAM, so 32 GB of RAM covers Q4_K_M. Speed depends on the hardware and is not quoted here because this site has not measured it. How the build names trade size for quality is covered in quantization explained, and the Ollama RAM and VRAM table lists other models by size.

Licence: what you may use it for

Codestral 22B is not open source. The model card on Hugging Face carries the licence tag mnpl and links to the Mistral AI Non-Production Licence, version 0.1. The repository was last updated in July 2025 and still carries that licence.

Mistral's announcement says the licence “means that you can use it for research and testing purposes” and that “commercial licenses are also available on demand by reaching out to the team”. The licence text also states that the model's outputs are not treated as derivatives of the model. This is a summary, not legal advice; read the licence before relying on it.

An earlier version of this page said Codestral had been relicensed under Apache 2.0. That could not be confirmed against Mistral's model card or licence file and has been removed. If you need a permissively licensed code model, see the alternatives below.

Model specifications

Fill-in-the-middle is what sets a completion model apart from a chat model. The editor sends the code before the cursor and the code after it, and the model writes what belongs in between. That is the mode an autocomplete plug-in uses; instruct mode is for asking questions about code or requesting a function.

Installation

The quickest route is Ollama. codestral, codestral:22b and codestral:v0.1 are the same 13 GB build. All 17 tags on ollama.com/library/codestral/tags list a 32K context.

# Default build, 13 GB
ollama pull codestral:22b
ollama run codestral:22b

# Smaller build for a 12 GB card, 8.3 GB
ollama pull codestral:22b-v0.1-q2_K

# Higher-precision build for a 24 GB card, 18 GB
ollama pull codestral:22b-v0.1-q6_K

Original weights with mistral-inference

Mistral recommends its own mistral-inference package for the unquantised weights. These two commands are from the model card:

pip install mistral_inference
mistral-chat $HOME/mistral_models/Codestral-22B-v0.1 --instruct --max_tokens 256

The BF16 weights are about 44 GB (22.25B parameters × 2 bytes), so that route needs a 48 GB card or two 24 GB cards. For editor autocomplete, the Continue.dev with Ollama guide covers pointing a VS Code or JetBrains plug-in at a local model, and best local autocomplete models compares the options.

What Mistral published about performance

Mistral's announcement evaluates Codestral on HumanEval and MBPP for Python generation, CruxEval for predicting program output, RepoBench for long-range repository-level completion, Spider for SQL, and HumanEval in six further languages. Its fill-in-the-middle results are compared against DeepSeek Coder 33B in Python, JavaScript and Java. The scores are published as chart images on that page, so they are not reproduced here; read them at the source.

The claim Mistral makes in words is about context: “With its larger context window of 32k (compared to 4k, 8k or 16k for competitors), Codestral outperforms all other models in RepoBench”. That was a statement about the models available in May 2024.

The one independent figure on the page is from a JetBrains researcher, who reported a Kotlin-HumanEval pass rate of 73.75 for Codestral at temperature 0.2, against 72.05 for GPT-4-Turbo and 54.66 for GPT-3.5-Turbo.

What has replaced it

Codestral 22B is the first Codestral and the only full-size one with downloadable weights. Mistral has since announced Codestral 25.01 (13 January 2025), which it describes as “about 2 times faster” than the original, and Codestral 25.08 (30 July 2025). Neither appears among Mistral's repositories on Hugging Face, where the only Codestral entries are this model and the 7B Mamba-Codestral, so they are hosted models rather than something you can pull.

For local use in 2026 the practical replacements are newer open-weight code models with permissive licences:

• Devstral — Mistral's own later code model. Devstral Small is Apache-2.0, 23.6B parameters, and the Ollama devstral:24b tag is 14 GB with a 128K context.
• Qwen 2.5 Coder 32B — Apache-2.0, for a 24 GB card.
• Qwen3-Coder-Next and the best local AI coding models list for the current picks.

Codestral 22B is still a reasonable choice for personal autocomplete on a 16 GB card, where its fill-in-the-middle mode is the point. For anything connected to paid work, the licence rules it out before quality comes into it.

Frequently asked questions

Sources

• Mistral AI: Codestral announcement — release date, context, languages, licence statement
• Codestral-22B-v0.1 model card — parameter count, usage modes, limitations
• Mistral AI Non-Production Licence 0.1 — the licence text
• Ollama tag list — every tag with its download size
• bartowski GGUF files — the file sizes in the hardware table