Codestral 22B vs StarCoder2 15B Instruct

Codestral 22B and StarCoder2 15B Instruct are close enough in size to make this a genuine architecture-and-training debate rather than a raw-parameter story. Codestral is Mistral's dedicated code model; StarCoder2 15B comes from the BigCode collaboration, trained on The Stack v2 with explicit permissive licensing (BigCode OpenRAIL-M). If open, redistributable weights are a hard requirement, StarCoder2's license is cleaner for many enterprise legal reviews.

On HumanEval, Codestral 22B scores roughly 81%, while StarCoder2 15B Instruct lands around 72–73%. The 7-8 point gap narrows on multi-language benchmarks where StarCoder2's broad training corpus — over 600 programming languages — gives it solid coverage of niche languages like Fortran, COBOL, or Elixir that Codestral may handle less gracefully.

For polyglot codebases with a mix of mainstream and legacy languages — say, a financial system touching Python, Scala, and COBOL — [StarCoder2 15B Instruct](/models/bigcode--starcoder2-15b-instruct) is worth testing. The breadth of training data may outweigh the raw benchmark gap on your actual distribution of code.

Codestral 22B is the better choice for pure Python/JavaScript/TypeScript-heavy workloads where HumanEval-style accuracy matters and you can absorb slightly higher inference cost. Its 32K context window also beats StarCoder2 15B's shorter window for file-level refactoring tasks. Check provider pricing on [Codestral 22B's model page](/models/mistralai--codestral-22b).

**Pick StarCoder2 15B Instruct** for polyglot or license-sensitive deployments. **Pick Codestral 22B** for mainstream-language accuracy and longer context needs.