China AI startup Moonshot AI unveiled Kimi-K2.6, the latest version in its open-source large language model Kimi series, SiliconANGLE reported on Monday.

Moonshot AI says it outperforms GPT-5.4 and Claude Opus 4.6 across several AI benchmarks.

Kimi-K2.6 is based on the SwiGLU activation function, which improves hardware efficiency and simplifies the LLM training process. The algorithm is also integrated into several open-source LLMs, including Meta's Llama series.

The Kimi-K2.6 model architecture consists of 384 "experts". Each expert is a small neural network optimised for a specific task. When it receives a prompt, only eight experts are used to generate a response, reducing hardware usage. It identifies the most important parts of a prompt using multihead latent attention (MLA) technology, which is more hardware-efficient than existing attention mechanisms.

It also includes a vision encoder with 400 million parameters, enabling it to process multimedia inputs as well as text. Based on this, Moonshot AI said it can create a complete website using only user instructions and interface sketches.

To handle complex tasks that take a long time, it runs up to 300 agents simultaneously. It splits an overall task into multiple pieces, and the agents process each piece in parallel to finish faster than sequential execution.

It also offers a "claw groups" feature in which humans and agents divide tasks. Low-level programming capabilities, such as in the Rust language, have also been improved from the previous version.