کیمی K2.6

Moonshot AI's open-source multimodal agentic model, advancing long-horizon coding, coding-driven design, proactive autonomous execution and swarm-based task orchestration. It is a 1T-parameter Mixture-of-Experts with 32B activated parameters and a vision encoder, supporting both visual and text input across a 256K context. Thinking is the default mode and can be switched off for an instant-response mode.

K2.6 is the release where Moonshot's argument moves from single tasks to sessions that run for most of a working day, and two of its own examples set the scale. Asked to download and run a small language model locally on a Mac, it implemented and optimised inference in Zig, a deliberately niche choice, across more than 4,000 tool calls, over twelve hours of continuous execution and fourteen iterations, taking throughput from about 15 to about 193 tokens per second - which the lab measures as roughly 20% faster than a widely used desktop runtime. In a second run it spent thirteen hours on an eight-year-old open-source financial matching engine, worked through twelve optimisation strategies and more than a thousand tool calls, read CPU and allocation flame graphs to find the real bottleneck, reconfigured the core thread topology, and reports throughput gains of 185% and 133% on the engine's two headline measures.

The other new capability is orchestration. Moonshot says the agent swarm now scales to 300 sub-agents across 4,000 coordinated steps, against 100 and 1,500 in the K2.5 research preview, and that it can turn a supplied document, spreadsheet or deck into a reusable skill that keeps the original's structure and style. Its own reliability team ran a K2.6-backed agent autonomously for five days handling monitoring, incident response and system operations, which is the kind of claim that is easy to state and hard to fake.

On benchmarks Moonshot reports 80.2 on SWE-bench Verified, 76.7 on the multilingual set, 58.6 on SWE-bench Pro, 66.7 on Terminal-Bench 2.0 and 89.6 on LiveCodeBench v6, with 90.5 on GPQA-Diamond and 92.5 F1 on DeepSearchQA. Coding-driven design is the other emphasis: complete front-end interfaces from a single prompt with deliberate layout, interaction and scroll-triggered animation, and simple full-stack flows spanning authentication, user interaction and database work.

The same table is where the limits sit, and they are worth reading. Moonshot's own numbers put K2.6 behind GPT-5.4 and Claude Opus 4.6 on several agentic sets, at 27.9 against 33.3 and 33.0 on APEX-Agents, 50.0 against 54.6 on Toolathlon and 55.9 against 62.5 on MCPMark, and at 34.7 on Humanity's Last Exam without tools it is behind every competing model in its own table. Broad reasoning is not what this release was aimed at.

One caveat is unusual enough to be worth repeating: because the weights are open, how the model is served changes what you get. Moonshot notes that reproducing its published numbers requires the official API and points at its own vendor-verification project for judging third-party hosts. That is the honest footnote attached to every open-weight release - the weights are the same everywhere, the serving is not. The copy served here is Moonshot's.