AI Briefing1T MoE Agent
Key point
Moonshot AI unveiled Kimi K2.6, a 1T MoE multimodal agent model.
Details
Moonshot AI unveiled Kimi K2.6. It's an open-source native multimodal agentic model, positioning long-horizon coding, design generation, autonomous execution, and swarm-style task orchestration as its core capabilities.
The model specs are 1T total parameters, 32B active, 61 layers, 384 experts, 8 experts selected per token, 256K context, and a MoonViT 400M vision encoder.
On benchmarks, it posts strong numbers in agentic tasks, coding, and vision.
- HLE-Full with tools 54.0
- BrowseComp 83.2, Agent Swarm 86.3
- DeepSearchQA accuracy 83.0, f1 92.5
- SWE-Bench Verified 80.2, SWE-Bench Pro 58.6
- LiveCodeBench v6 89.6, MMMU-Pro 79.4
It's compared against GPT-5.4, Claude Opus 4.6, Gemini 3.1 Pro, and Kimi K2.5, showing top-tier performance on some agentic and coding items. The release page also highlights 300 sub-agents, swarm orchestration at the level of about 4,000 steps, and 24/7 background agent scenarios.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.