Codestral 25.01

Mistral AI identifies Codestral 25.01 as an evolution of the previous 22B model with a 'more efficient architecture' and an improved tokenizer. However, specific architectural modifications beyond the context window expansion (to 256k) and the 'dense' nature of the model are not publicly documented in a technical paper or detailed model card. The lack of a formal technical report for this specific version leaves the exact training methodology and architectural changes opaque.

Information regarding the training data is extremely limited. Mistral AI only provides vague marketing claims stating the model was trained on 'diverse codebases' and supports '80+ programming languages.' There is no public disclosure of specific data sources, dataset proportions, filtering techniques, or cleaning methodologies. The absence of a data provenance report or sample data availability results in a low score.

Mistral mentions an 'improved tokenizer' that contributes to a 2x speed increase compared to the previous version. While the tokenizer is accessible via the 'mistral-common' library and integrated into platforms like Hugging Face, detailed documentation on the vocabulary size or the specific training data alignment for this version is not explicitly provided in the primary release notes.

Despite being a major release, Mistral AI has not officially disclosed the exact parameter count for Codestral 25.01. While third-party sources and previous versions suggest it is in the 'sub-100B' or '22B-24B' range, the official documentation remains silent on the specific density. This lack of clarity on a fundamental model metric is a significant transparency gap.

There is zero public information regarding the compute resources used to train Codestral 25.01. No disclosure of GPU/TPU hours, hardware specifications, training duration, or environmental impact (carbon footprint) has been made available by Mistral AI.

Mistral provides a range of benchmark results (HumanEval, MBPP, CruxEval, etc.) in their announcement blog post. However, they do not release the exact evaluation code, specific prompts, or few-shot examples required to reproduce these results exactly. While the model is tracked on public leaderboards like LMSYS Copilot Arena, the internal benchmarking methodology remains largely proprietary.

The model consistently identifies itself as a Mistral AI product and is correctly versioned as '25.01' in API endpoints and documentation. It maintains a clear identity as a specialized coding model and does not exhibit the identity confusion seen in some other models that claim to be competitors.

The licensing for Codestral 25.01 is restrictive. While the weights are available for download, they are governed by the 'Mistral AI Non-Production License,' which prohibits commercial use without a separate agreement. This 'open-weights but not open-source' approach creates ambiguity for developers, as it does not meet the standard definition of Open Source (e.g., Apache 2.0).

Basic guidance on running the model is available through community integrations (like Ollama and Unsloth), and VRAM requirements for various quantization levels are documented by third-party tools. However, Mistral's official documentation lacks a comprehensive hardware requirement guide, particularly regarding memory scaling at the maximum 256k context length.

Mistral uses a clear date-based versioning system (25.01), which is an improvement over vague naming conventions. However, there is no public changelog or detailed documentation of behavioral drift or 'alignment tax' between this version and its predecessor. Updates to the 'codestral-latest' API alias can lead to silent behavior changes for users not pinning specific versions.