Kimi K2.6 Released - Progress in Open-Source Coding

Kimi has released K2.6. The core improvements span long-horizon coding, agentic tasks, frontend/full-stack generation, and multi-agent orchestration overall.

• Long-horizon coding performance: Generalization has improved across Rust, Go, Python, frontend, devops, and performance optimization, showing a large improvement over K2.5 on the internal Kimi Code Bench.
• Continuously executing tasks: Through thousands of tool calls and over 12 hours of continuous execution, it significantly boosted throughput in Zig-based inference optimization and exchange-core overhauls.
• Real-world optimization case: Restructured exchange-core over 13 hours, with over 1,000 tool calls and over 4,000 lines modified, achieving a 185% improvement in median throughput and a 133% improvement in performance throughput.
• Frontend/full-stack: Turns simple prompts into polished UIs, leverages image and video generation tools, and supports lightweight full-stack workflows including authentication and databases.
• Agent Swarm expansion: With 300 sub-agents and 4,000 coordination steps, it performs parallel search, research, document writing, and file creation, and the styles of PDFs, slides, spreadsheets, and Word documents can be reused as skills.
• Proactive agent: Handles long-duration autonomous operation, multi-agent collaboration, and even task reassignment, and states that tool-calling quality and long-run reliability have improved on the internal Claw Bench and in enterprise beta testing.
• Benchmarks: Presented key figures including SWE-Bench Verified 80.2, LiveCodeBench v6 89.6, BrowseComp (agent swarm) 86.3, and AIME 2026 96.4.

Benchmark reproduction via the official API and vendor-verified tools is recommended, and K2.6 is positioned not as a simple coding model but as an agent model strong in long-duration autonomous execution and tool use.