Moonshot AI · 2026-04-13 · notable
Kimi Code K2.6 — Moonshot's Upgraded Coding Agent Model
Moonshot rolls out K2.6-code-preview to all Kimi Code subscribers — deeper reasoning traces, cleaner multi-step agent plans, and more reliable tool execution. Early testers called it 'Opus-flavored.'
Moonshot's Kimi Code gets a major model upgrade with deeper reasoning and cleaner agent planning — subscribers say it feels like Opus.
Key specs
| Context window | 256K tokens |
|---|---|
| Architecture | 1T total / 32B active MoE |
| K2.5 swe bench verified | 76.8% |
| K2.5 live code bench | 85% |
What is it?
Kimi Code K2.6-code-preview is the latest coding-specialized model from Moonshot AI, powering the Kimi Code terminal-based coding agent. It rolled out to all Kimi Code subscribers on April 13 after a week of closed beta testing. The model is built on Moonshot's 1-trillion-parameter MoE architecture with 32B active parameters per forward pass.
How does it work?
K2.6 produces deeper reasoning traces than K2.5, with more structured thinking patterns similar to Claude Opus's verbose chain-of-thought approach. The model handles multi-step tool calls more reliably and decomposes complex tasks into parallelizable sub-tasks that the Agent Swarm infrastructure can distribute across up to 100 specialized sub-agents. The CLI update (v1.33.0) that accompanied the rollout dropped hardcoded K2.5 references.
Why does it matter?
Moonshot's Kimi K2.5 already matched frontier models on SWE-Bench Verified (76.8%) at a fraction of the price. Early K2.6 testers on Reddit described the reasoning upgrade as 'Opus-flavored' — a notable comparison given that Claude Opus is the benchmark for verbose, reliable reasoning. If the jump from K2.5 to K2.6 mirrors the K2 to K2-Thinking leap, this is a meaningful step up.
Who is it for?
Developers using Kimi Code for terminal-based coding assistance.
Try it
curl -L code.kimi.com/install.sh | bash