Key takeaways

• Mistral Large 3 (2512) is an open-weight, general-purpose multimodal model using a granular Mixture-of-Experts (MoE) architecture with 41B active and 675B total parameters. It's designed for instruction following and mathematical problem-solving under an Apache 2.0 license.
• The model demonstrates strong capabilities in instruction following, generalization, and mathematical reasoning, achieving 94.97% accuracy on the 'MATH-500' dataset. It excels in diverse mathematical categories with detailed, step-by-step reasoning.
• Mistral Large 3 (2512) struggles with complex trigonometric/geometry problems, intricate logical deduction rules, and situations requiring meticulous calculation verification or nuanced language interpretation to avoid miscalculations or incorrect problem setup.
• The model's worst performance was 5.07% accuracy (1 out of 58 questions) on the 'Humanity's Last Exam' benchmark, struggling with multi-step reasoning, precise rule application, complex calculations, and specific expert-level factual recall across various academic subjects.
• Common failure modes include misinterpreting problem parameters, arithmetic errors in multi-step calculations, incorrect application of mathematical theorems (especially in geometry), and ambiguities in problem statements leading to incorrect assumptions or derived results.

Technical specifications

Context length

262k tokens

Max output

—

Architecture

MoE

Modalities

text+image->text

Parameters

~675B

Open weights

Yes

Licence

Apache 2.0

Regions

France

Vendor

MistralAI

Provider model key

mistralai/mistral-large-2512