Key takeaways

• Moonshot AI's Kimi K2.6 is an intelligent multimodal model (text+image->text) excelling in general intelligent tasks, Agent tasks, coding, and visual understanding, designed for both dialogue and autonomous agent execution.
• It features a Mixture-of-Experts (MoE) architecture with 1 billion parameters, released with open weights under an Apache 2.0 license, promoting accessibility and research.
• Kimi K2.6 demonstrates strong mathematical reasoning with 96.8% accuracy on the MATH-500 dataset, particularly in structured problem-solving across Algebra, Number Theory, and Geometry, although it sometimes struggles with strict output formatting.
• The model achieves only 33.33% accuracy in Text-to-SQL generation (BIRD-CRITIC dataset), showing proficiency in complex JSONB and advanced window functions but significantly struggling with recursive CTEs, complex date/time conditions, and ON CONFLICT clauses.
• Its high accuracy in math tasks represents a best-case for structured logical reasoning, while its low accuracy in multi-turn Text-to-SQL highlights a significant limitation in understanding complex SQL constructs and conditional reasoning.

Technical specifications

Context length

256k tokens

Max output

—

Architecture

MoE

Modalities

text+image->text

Parameters

~1000B

Open weights

Yes

Licence

Apache 2.0

Regions

China

Vendor

Moonshot AI

Provider model key

moonshot/kimi-k2.6