Key takeaways

• Codestral (2508) is a Transformer-based model specializing in low-latency, high-frequency coding tasks like fill-in-the-middle, code correction, and test generation.
• The model excels in generating Python code for well-defined tasks, string manipulation, and leveraging Python libraries, achieving a 50% pass rate on LiveCodeBench for such problems.
• Despite its coding strengths, Codestral struggles significantly with intricate logic, complex conditional problems, and competitive programming challenges, showing a low 10% pass rate on Terminal-Bench and frequent incorrect solutions.
• In mathematical problem-solving, it shows solid performance in fundamental problems (Algebra, basic Number Theory) but its accuracy sharply declines with increasing complexity, particularly in Geometry and Precalculus.
• The model achieved a very low accuracy of 6.67% on advanced mathematical problems like AIME 2025, solving only 1 out of 30 problems correctly, highlighting difficulties with complex combinatorial, geometric, and precise arithmetic challenges.

Technical specifications

Context length

256k tokens

Max output

—

Architecture

Transformer

Modalities

text->text

Parameters

—

Open weights

No

Licence

Apache 2.0

Regions

France

Vendor

MistralAI

Provider model key

mistralai/codestral-2508