Key takeaways

• The Kimi K2 Thinking Turbo model excels in complex mathematical reasoning, achieving 94.1% accuracy on the MATH-500 dataset by providing detailed, step-by-step solutions across Algebra, Geometry, and Number Theory. Its main purpose is high-speed, long-context text generation with advanced reasoning.
• It is an open-weights Mixture-of-Experts (MoE) model. This architecture enhances efficient processing and suggests strong performance in specialized domains. It supports a 256K context window and offers output speeds of 60-100 tokens per second.
• While strong in mathematical reasoning, the model sometimes struggles with precise output formatting (e.g., interval notation) and specific geometric interpretations. This can lead to minor errors in exact-match scoring despite conceptual correctness.
• A significant limitation is its poor performance (6.25% accuracy) on the Terminal-Bench (Terminus-1) dataset, particularly in complex programming, system administration, and multi-step execution tasks. It often returns empty or unattempted responses in these scenarios.
• Common failure modes include incorrect output formatting, misinterpretation of geometric properties or algebraic errors in math, and a general inability to handle precise instructions and generate accurate code/commands for programming tasks.

Technical specifications

Context length

262k tokens

Max output

—

Architecture

MoE

Modalities

text->text

Parameters

~1000B

Open weights

Yes

Licence

Open

Regions

China

Vendor

Moonshot AI

Provider model key

moonshot_kimi-k2-thinking-turbo