Kimi K2.6
#20moonshot-ai · seit 2026-04-20 · 131× · zuletzt 30. Juni 2026
Kimi K2.6 is an open-weight frontier language model released by Moonshot AI on April 20, 2026, featuring 1 trillion total parameters via a Mixture-of-Experts architecture with 32B parameters activated per token. It targets long-horizon autonomous coding, multimodal input (text, image, video via native MoonViT encoder), and multi-agent orchestration scaling to 300 parallel sub-agents. Weights are freely available on Hugging Face under a Modified MIT License; API access is provided via platform.moonshot.ai with OpenAI/Anthropic-compatible endpoints. On agentic benchmarks such as SWE-Bench Pro (58.6%) and HLE-Full with tools (54.0%), the model reportedly outperforms GPT-5.4 and Claude Opus 4.6 per vendor-reported numbers.
Fonctionnalités
| Context Window (Tokens) | 262,144 tokens (262K) |
| License | Modified MIT License (free for commercial use; branding 'Kimi K2' in UI required if >100 million MAU or >$20 million USD monthly revenue) |
| Multimodality | Natively multimodal: text, images, and video (mp4, mov, avi, webm, and others, recommended up to 2K resolution) via integrated MoonViT-3D vision encoder (400 million parameters); no separate module/adapter |
| Release Date | April 20, 2026 (GA; code preview from April 13, 2026) |
