Kimi K2.6
Kimi K2.6 is a 1T-parameter MoE that holds coding tasks together for up to 12 hours with 4,000 coordinated steps and supports 300 parallel sub-agents per swarm.
Moonshot · Kimi
Kimi K2.6
Agentic-coding flagship with 12-hour task horizons
Kimi K2.6 is a 1T-parameter MoE that holds coding tasks together for up to 12 hours with 4,000 coordinated steps and supports 300 parallel sub-agents per swarm.
Key features
Kimi K2.6 · AI Models
Parameters
1T total · 32B active
Context window
262K tokens
Released
April 20, 2026
Pricing
Open weights · Kimi API
Key features
- 1T total / 32B active parameters · 384 experts (8 active).
- 12-hour autonomous coding sessions with 4,000 coordinated steps.
- Up to 300 parallel sub-agents per swarm.
- 80.2% on SWE-Bench Verified, 93.2% on MathVision.
Best for
Pick Kimi K2.6 when you need a long-running autonomous coding agent — overnight refactors, multi-day research sessions, swarm orchestration.
Frequently Asked Questions
Is Kimi K2.6 open-source?
Yes — Moonshot AI publishes the K2.6 weights on Hugging Face.
How does Kimi K2.6 compare to Claude Opus 4.7?
K2.6 is open-weight and built specifically for very long autonomous coding sessions, while Claude Opus 4.7 is a closed flagship with broader general capabilities and a 1M-token context.
Open Chat
Kimi K2.6 is a 1T-parameter MoE that holds coding tasks together for up to 12 hours with 4,000 coordinated steps and supports 300 parallel sub-agents per swarm.