Key takeaways

• Mistral Medium 3.5 is a 128B, Transformer-based instruction-following model designed for agentic workflows, coding, and complex reasoning tasks, supporting both text and image inputs with text output.
• The model boasts a very large 256k context window and unifies instruction-following, reasoning, and coding into a single checkpoint. It is developed by Mistral AI, released with a Modified MIT license, and has open weights.
• It demonstrates strong mathematical problem-solving capabilities, particularly in algebra and number theory, and good logical reasoning in some Math problems, successfully breaking down complex equations or game theory scenarios.
• Key limitations include struggles with precise numerical calculations, intricate logical deductions from complex statements (especially with visual interpretations), and its performance decreases significantly with increasing problem difficulty across all categories.
• On the MATH-500 dataset, it achieved 93.6% accuracy, showing strengths in algebra and mathematical formulas. However, on the "Humanity's Last Exam" benchmark, it scored only 4.44% accuracy, struggling with multi-modal tasks, specific factual recall for niche knowledge, and complex computational constraints.

Technical specifications

Context length

262k tokens

Max output

—

Architecture

Transformer

Modalities

text+image->text

Parameters

~128B

Open weights

Yes

Licence

Modified MIT

Regions

France

Vendor

MistralAI

Provider model key

mistralai/mistral-medium-3-5