Moonshot AI · 2026-04-20 · seismic
Kimi K2.6 — Moonshot AI's 1T-Parameter Agentic MoE Opens Under Modified MIT
Moonshot AI releases Kimi K2.6 with open weights — a 1T-parameter MoE model with 32B active params, 256K context, and native vision via MoonViT. Achieves 58.6% on SWE-Bench Pro and supports agent swarms up to 300 sub-agents. Modified MIT license.
Moonshot AI open-sources a 1-trillion-parameter multimodal model built for orchestrating swarms of up to 300 parallel sub-agents.
Key specs
| Active params | 32B |
|---|---|
| Context window | 256K tokens |
| Total parameters | 1T |
| Swe bench pro | 58.6% |
| Aime 2026 | 96.4% |
What is it?
Kimi K2.6 is Moonshot AI's latest open-weights model — a 1T total / 32B active parameter Mixture-of-Experts architecture with 384 experts, 256K-token context, and a built-in 400M-parameter vision encoder called MoonViT. Available under a Modified MIT license on HuggingFace, it supports both 'thinking' (extended reasoning) and 'instant' modes for text, images, and video.
How does it work?
The model uses Moonshot's Hybrid Attention design with Compressed Sparse Attention and Heavily Compressed Attention layers, reducing KV cache memory at long contexts. Agent swarm support is a first-class feature: the model is designed to coordinate fleets of up to 300 parallel sub-agents executing up to 4,000 coordinated steps with native long-horizon task decomposition. MoonViT handles images and video natively in the same context window as text.
Why does it matter?
At 58.6% on SWE-Bench Pro and 96.4% on AIME 2026 with open weights under a permissive license, this raises the self-hostable coding agent bar. The built-in agent swarm coordination makes it one of the first open models explicitly designed for multi-agent pipeline work at scale.
Who is it for?
ML practitioners, self-hosters, teams running agentic coding pipelines
Try it
huggingface.co/moonshotai/Kimi-K2.6