Moonshot AI Releases Kimi K2.6: Open-Source Giant with 1T Parameters and 300-Agent Swarms

Moonshot AI has released Kimi K2.6, the latest iteration of its Kimi open-source model series, marking a significant leap in what open-weight models can accomplish for agentic and coding tasks. Released on April 20, 2026, K2.6 is available across Kimi.com, the Kimi App, API, and as downloadable weights on Hugging Face under a Modified MIT License.

The model uses a Mixture-of-Experts architecture with 1 trillion total parameters, activating 32 billion per token. It employs 384 experts with 8 activated per prompt, a 256K token context window, and a 400-million-parameter MoonViT vision encoder that processes both images and video natively. Multi-head Latent Attention (MLA) reduces memory requirements without sacrificing performance, and a SwiGLU activation function improves training stability and hardware efficiency across 61 transformer layers.

The headline capability is Kimi K2.6's agent swarm architecture, which allows the model to deploy up to 300 parallel sub-agents executing across 4,000 coordinated steps simultaneously — a threefold improvement over K2.5's limit of 100 agents. In one published case study, the model autonomously optimized a financial matching engine over 13 hours, achieving a 185% medium throughput increase and 133% performance gain through systematic code modifications with minimal human oversight.

On the HLE-Full benchmark — one of the most demanding agentic evaluations — Kimi K2.6 scores 54.0, edging out GPT-5.4 at 52.1 and Claude Opus 4.6 at 53.0. It also scores 58.6 on SWE-Bench Pro, which measures real-world GitHub issue resolution. Moonshot says the model shows particular strength in Rust development and front-end generation from natural language descriptions.

For deployment, the model runs on vLLM, SGLang, or KTransformers with two inference modes: Thinking mode for complex multi-step reasoning and Instant mode for lower-latency interactive use. A new "claw groups" feature enables seamless human-AI task handoffs within the same agentic loop. With open weights and competitive benchmark performance, Kimi K2.6 cements Moonshot AI's position as a major challenger to closed frontier models from OpenAI and Anthropic.