Moonshot AI today released Kimi-K2.6, the latest addition to its popular Kimi series of open-source large language models.

The Chinese artificial intelligence startup says that the algorithm outperforms GPT-5.4 and Claude Opus 4.6 across several AI benchmarks.

Each of an LLM’s artificial neurons includes modules called weights that prioritize input data based on relevance. From there, the data is sent to an algorithm called an activation function. It processes the input and then determines whether the output is valuable enough to be shared with other neurons. 

Kimi-K2.6 implements an activation function known as the Swish-Gated Linear Unit, or SwiGLU for short. It’s more hardware-efficient than earlier algorithms and simplifies the LLM training process in certain respects. The algorithm has been integrated into several other open-source LLM families besides Kimi, most notably Meta Platforms Inc.’s Llama series.

Kimi-K2.6’s neurons are organized into 384 so-called experts, miniature neural networks that are each optimized for a different set of tasks. When the LLM receives a prompt, it only uses eight experts to generate a response. Reducing the number of neural networks involved in processing user input lowers hardware usage.

Kimi-K2.6’s neural networks use a technology called MLA, or multi-head latent attention, to identify the most important part of a prompt. It’s a more hardware-efficient version of the standard attention mechanism found in LLMs. The technology works the same way, except that it compresses the data it processes into a lightweight mathematical representation to reduce hardware requirements.

Kimi-K2.6’s neural networks are supported by a vision encoder with 400 million parameters. It turns images into embeddings, mathematical representations that the LLM can more easily understand. The vision encoder enables Kimi-K2.6 to process not only text prompts but also multimedia input.

According to Moonshot AI, the model can turn simple user instructions and interface sketches into complete websites. When the LLM is given a more complex, time-consuming task, it can launch as many as 300 agents to speed up the workflow. The agents break down a task into substeps and perform them in parallel, which is faster than completing them one after one another.

Kimi-K2.6 can optionally loop in human workers using a feature called claw groups. According to Moonshot AI, it enables the LLM to split the work involved in a project between humans and agents. Kimi-K.26 is also better than its predecessor at certain other tasks, including Rust development. Rust is a low-level language with a complex syntax that is mainly used to program connected devices.

Moonshot AI compared Kim-K.26 to GPT-5.4 and Claude Opus 4.6 across more than two dozen popular benchmarks. According to the company, its algorithm bested the two frontier LLMs or came within a few percentage points of their scores in most tests. 

One of the evaluations that Kim-K.26 completed more effectively is HLE-Full, which ranks among the AI ecosystem’s most difficult benchmarks. It comprises about 2,500 doctorate-level questions spanning more than 100 academic fields. Kim-K.26 scored 54 while Opus 4.6 and GPT 5.4 earned 53 and 52.1 points, respectively.

Image: Unsplash