Back to Gap Map Model components / Fine-tuned / chat models

Codestral

restricted / Overall score: 3.0

Mistral's code-completion and fill-in-the-middle specialist, instruction-tuned with a chat mode and reached through Mistral's API, a customer VPC or on-premises deployment. It sits in a broader coding stack alongside Codestral Embed for retrieval, Devstral as the open-weight agentic counterpart, and the Mistral Vibe CLI.

A coding specialist rather than a general chat model. Mistral published this model's original benchmark scores as chart images rather than text.

Openness

2 high confidence

2.0

weights: open(downloadable on HF)
data: closed
code: closed(no training pipeline)
license: Mistral-AI-Non-Production-License(MNPL, non-OSI, research/testing only

Weights are downloadable, but only under the Mistral AI Non-Production License: not OSI-approved and limited to research and testing, with a commercial license available on request. Open weights behind a use-restricting license, so restricted rather than open weights.

https://huggingface.co/mistralai/Codestral-22B-v0.1/raw/main/README.md recorded 2026-09-29
Model card front matter "license_name: mnpl" and body "Codestral-22B-v0.1 is released under the `MNLP-0.1` license." Searched the card for training data, dataset, training code and pipeline: none published.
https://mistral.ai/licenses/MNPL-0.1.md recorded 2026-09-29
License text: "You shall not supply the Mistral Models or Derivatives in the course of a commercial activity" and use is limited to "testing, research, Personal, or evaluation purposes in Non-Production Environments".
https://mistral.ai/news/codestral/ recorded 2026-09-29
Launch post: "a 22B open-weight model licensed under the new" Mistral AI Non-Production License, "which means that you can use it for research and testing purposes" and "Codestral can be downloaded on" Hugging Face. Searched the body for training data, dataset and source code releases: none.
https://mistral.ai/news/codestral/ recorded 2026-09-29
Launch post: "Commercial licenses are also available on demand"

Adoption

3 low confidence

3.0

Codestral is integrated into IDE assistants such as Continue and Tabnine, into LlamaIndex and LangChain, and into Mistral's own coding stack. Mistral reports version-over-version gains in completion acceptance but publishes no download or user count, so this reading reflects how widely the model is embedded rather than a verified figure.

https://mistral.ai/news/codestral-25-08/ recorded 2026-08-13
"+30% increase in accepted completions, +10% more retained code after suggestion, 50% fewer runaway generations"; distributed via API, VPC and on-prem; no download or user count published

Capability

3 medium confidence

3.0

Codestral is a strong code-completion specialist for its size, but it is a narrow fill-in-the-middle model. No SWE-bench result for Codestral appears on any Mistral page, so its placement below current agentic coding models is based on the completion and fill-in-the-middle benchmarks.

https://mistral.ai/news/codestral-2501 recorded 2026-09-29
Codestral 25.01 post: "We have benchmarked the new Codestral with the leading sub-100B parameter coding models"; tables give Codestral-2501 (256k context, HumanEval "86.6%", MBPP "80.2%", RepoBench "38.0%", HumanEvalFIM average "85.9%") and Codestral-2405 22B (32k context, HumanEval "81.1%", MBPP "78.2%"). Searched the body for SWE-bench: none.

Verified 2026-09-29, except some scores, still awaiting re-confirmation