Kimi K2.6: The Evolution of Open-Source Coding

Moonshot AI has open-sourced Kimi K2.6. The core highlights are state-of-the-art coding, long-horizon execution, and agent swarm capabilities. The model is available on Kimi.com, Kimi App, API, and Kimi Code.

The performance overview emphasizes benchmark improvements across general agents, coding, and visual agents.

• General Agents: Humanity's Last Exam (Full) w/ tools, BrowseComp, DeepSearchQA, Toolathlon, OSWorld-Verified
• Coding: Terminal-Bench 2.0, SWE-Bench Pro, SWE-Multilingual
• Visual Agents: MathVision w/ python, V w/ python*

A long-horizon coding case was also presented.

• On a Mac, Qwen3.5-0.8B was deployed locally, and inference was optimized with Zig, performing 4,000+ tool calls over 12+ hours across 14 iterations.
• Throughput improved from roughly 15 tok/s to 193 tok/s, ultimately reportedly achieving about 20% faster speed than LM Studio.
• The 8-year-old open-source financial matching engine exchange-core was refactored over 13 hours, involving 1,000+ tool calls and 4,000+ lines of changes, applying 12 optimization strategies.
• As a result, it claimed a 185% increase in medium throughput (0.43 → 1.24 MT/s) and a 133% increase in performance throughput (1.23 → 2.86 MT/s).

It also introduced capabilities spanning frontend generation, animation, full-stack workflows, and use of image/video generation tools, demonstrated through internal benchmarks Kimi Code Bench and Kimi Design Bench. The design generation examples show, from a single prompt, the creation of a polished landing page, a simple full-stack app with auth and a DB, and a tool-integrated website.

Finally, it highlights the Agent Swarm feature. Rather than scaling up a single model, this structure decomposes tasks into multiple specialized agents that run in parallel, and it explains that K2.6 offers a more advanced swarm experience than K2.5.