Moonshot AI has launched Kimi-K2.6, the latest large language model in its Kimi series.
The company asserts that Kimi-K2.6 surpasses leading models like GPT-5.4 and Claude Opus 4.6 on multiple AI benchmarks.
Kimi-K2.6 features an efficient Swish-Gated Linear Unit (SwiGLU) activation function, streamlining the training process. This model organizes its neural networks into 384 specialized 'experts,' using only 8 per response to minimize hardware demands.
Utilizing multi-head latent attention (MLA), the model efficiently identifies key prompt elements by compressing data. It also incorporates a 400-million-parameter vision encoder, enabling it to process image inputs and generate code from sketches.
For complex tasks, Kimi-K2.6 can deploy up to 300 agents to work in parallel. The model also offers 'claw groups,' allowing it to integrate human workers into projects. Moonshot AI reports Kimi-K2.6 shows improved performance in areas like Rust development.
In performance tests, including the challenging HLE-Full benchmark, Kimi-K2.6 achieved a score of 54, slightly outperforming Opus 4.6 (53) and GPT 5.4 (52.1).