Moonshot AI
Kimi K2.6
The open-source favourite for agentic coding.
Open weightstextimageREL 2026-04-20
Compare this model →Verified 2026-06-12
Overview
Kimi K2.6 is Moonshot's open-weights MoE model, widely regarded as the best open option for agentic coding workflows. 262K context, vision input, $0.67/$3.39 per 1M via API — or self-hosted on serious hardware.
Benchmark scores
Methodology →Intelligence54
Coding80.2
Reasoning90.5
Math96.4
Sources and verification
- Math: AIME 2026 96.4
- Coding: SWE-bench Verified 80.2
- Intelligence: Artificial Analysis Intelligence Index · checked 2026-06-12
- Reasoning: GPQA Diamond 90.5
- Catalogue and pricing: OpenRouter model record · checked 2026-06-12
Strengths and weaknesses
Strengths
- Best open-weights coding agent
- Vision input
- Full 262K output window
Trade-offs
- Smaller context than 1M-class rivals
- Heavy to self-host (1T-class MoE)
Reach for this model when
Open-stack coding agents — Claude Code-style workflows without closed APIs.
Run something like it locally
Self-hostable via vLLM on multi-GPU servers; quantised community builds exist for high-VRAM workstations.
