Kimi K2.6

• 🆕 Moonshot AI's open-weight agentic model, released April 2026.
• 🧠 Mixture-of-Experts: 1 trillion total parameters, 32B active per token.
• 📏 256K-token context window with native INT4 quantization.
• 👁️ Native multimodal — accepts text, image, and video inputs.
• 🔧 Function calling and web search among built-in capabilities.
• 🎯 Tuned for long-horizon coding and agent orchestration.
• 💬 Successor to Kimi K2.5 within the Kimi K2 family.
• 🌐 Open weights distributed publicly on Hugging Face.

Moonshot

Moonshot is an AI research lab known for developing the Kimi family of large language models. The organization has gained recognition for building capable reasoning-oriented models, with the Kimi line representing its flagship series of text generation…

Kimi K2.6 is Moonshot AI's open-source, native multimodal agentic model, released on April 20, 2026 and distributed with open weights on Hugging Face. It uses a Mixture-of-Experts design with 1 trillion total parameters and roughly 32 billion activated per token, paired with a vision encoder so the model can take text, image, and video inputs natively rather than through a bolted-on adapter.

The model carries a 256K-token context window and ships in a native INT4 quantization, which keeps memory and serving footprint lower than a full-precision deployment. Its stated capabilities span reasoning, vision, code generation, function calling, and web search, and Moonshot positions it specifically toward long-horizon coding and proactive, autonomous task execution.

Kimi K2.6 is the direct successor to Kimi K2.5, released in January 2026, and continues the same Kimi K2 lineage. Moonshot frames the newer release around agentic work — long-horizon coding, coding-driven design, and agent swarm orchestration — emphasizing sustained execution across many coordinated steps rather than a change in raw model size.

Because independently verified benchmark figures for this release were not available among the trusted sources, this entry describes only its documented architecture, context length, modalities, and intended use rather than performance scores.