Kimi K2.6🔒Private

• 🏢 Moonshot AI's open-weight Kimi K2.6, served inside a Trusted Execution Environment.
• 🔒 Hardware attestation lets you verify enclave identity and configuration independently.
• 🧠 Trillion-parameter Mixture-of-Experts, roughly 32B parameters active per token.
• 📏 262,144-token context, the setting used in Moonshot's own evaluations.
• 👁️ Native multimodal input: text plus images, via a MoonViT vision encoder.
• 🔧 Built for long-horizon coding, tool calling and swarm-style agent orchestration.
• 🎯 Moonshot reports 36.4% on the HLE text-only subset without tools.

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-weight, natively multimodal agentic model, offered here in a confidential-computing configuration: the model runs inside a Trusted Execution Environment, and hardware attestation evidence is published so users can verify which weights and configuration the enclave is actually running. That makes it aimed at workloads where prompt and output confidentiality matter as much as capability.

Architecturally it is a Mixture-of-Experts design with about one trillion total parameters and roughly 32 billion active per token, paired with a MoonViT vision encoder that accepts images alongside text. Moonshot positions it for long-horizon coding, proactive autonomous execution and swarm-based task orchestration, with tool calling and structured outputs supported for agent frameworks. Evaluations in the model card were run at a 262,144-token context length.

It follows Kimi K2.5 in Moonshot's K2 line, with the vendor's published evaluations emphasising agentic and long-horizon behaviour. On Moonshot's own reported figures, K2.6 reaches 36.4% on the text-only subset of Humanity's Last Exam without tools and 55.5% with tools, with thinking mode enabled.

Related models in the wider Kimi lineup include Kimi K2.7 Code, which narrows the focus to coding, along with Kimi K3 and its latency-oriented variant Kimi K3 Fast. K2.6 is the choice when open weights, multimodal input and verifiable private execution are the priorities.