Key takeaways

• MistralAI's Magistral Small 1.2 (2509) is a Transformer-based model specializing in unsupervised text-to-text generation for mathematical reasoning, particularly lower algebraic levels.
• The model excels in basic algebraic tasks (Algebra_level_1, Algebra_level_2) on the MATH-500 dataset, achieving 95.4% accuracy with highly readable and well-formatted solutions.
• A significant limitation is its poor performance in complex tasks requiring coding, debugging, and environment interaction, scoring only 1.25% accuracy on Terminal-Bench (Terminus-1), often resulting in empty outputs.
• Magistral Small 1.2 demonstrates strong algebraic manipulation and produces clear math solutions but struggles with generating correct code or effectively debugging, exhibiting a near-zero success rate on Terminal-Bench.
• The model performs best on mathematical reasoning with high accuracy and readability for basic algebra, but its worst performance is on coding and system administration tasks, where it largely fails to provide solutions.

Technical specifications

Context length

131k tokens

Max output

—

Architecture

Transformer

Modalities

text->text

Parameters

~24B

Open weights

Yes

Licence

Proprietary

Regions

France

Vendor

MistralAI

Provider model key

mistralai/magistral-small-2509