Wenxiu AI, open-source LLM 'Kimi-K2.6' released... directly compared with GPT-5.4

Chinese AI startup Moonshot AI has unveiled its latest open-source large language model (LLM), “Kimi-K2.6.” The company states that this model either leads or is slightly behind GPT-5.4 and Claude Opus 4.6 in major AI benchmark tests.

Kimi-K2.6 is the newest addition to Moonshot AI’s “Kimi” series. This model is designed to handle not only text but also multimedia inputs such as images. Moonshot AI emphasizes its focus on efficiency and practical task execution. This means not only improved response quality but also optimized architecture to enable the model to perform more complex tasks using the same computational resources.

A structure that boosts performance with fewer resources is key

Kimi-K2.6 employs an activation function called “Swiglu.” This approach improves hardware utilization efficiency compared to traditional methods and partially simplifies the model training process. This method has also been applied in open-source models like Meta Platforms’ Llama series.

Inside the model, 384 “expert” networks are deployed. However, not all these networks run simultaneously for each user query. When generating responses, only 8 experts are selectively activated. This “expert mixture” approach activates only the necessary parts, reducing computational load and infrastructure requirements.

Additionally, it incorporates “multi-head latent attention” technology. This is a form of attention mechanism that more efficiently filters important information from prompts. Because it compresses data into lighter mathematical representations, it requires less hardware compared to standard attention structures.

Enhanced image understanding and agent collaboration features

Kimi-K2.6 is also equipped with a visual encoder with 4 billion parameters. This component converts images into “embeddings” that are easier for the model to understand. As a result, Kimi-K2.6 can process not only sentence inputs but also work with sketches or visual materials.

According to Moonshot AI, the model can generate complete websites based solely on simple user instructions and interface sketches. For more complex and time-consuming tasks, up to 300 intelligent agents can work collaboratively simultaneously. This approach involves each agent breaking down tasks into smaller sub-steps for parallel processing, which can speed up execution compared to sequential methods.

Furthermore, through a “group collaboration” feature, workflows can be designed for humans and AI agents to work together. Moonshot AI explains that this allows assigning work tasks within projects to both human workers and AI agents. The company adds that, compared to previous generations, its performance on high-difficulty programming tasks like Rust development has also improved.

Claims to outperform GPT-5.4 in high-difficulty evaluation HLE-Full

Moonshot AI states that Kimi-K2.6 has been compared with GPT-5.4 and Claude Opus 4.6 across more than 20 major benchmarks. The company claims that the new model outperforms these top models in several tests or narrows the gap to within a few percentage points.

Especially in the highly rigorous “HLE-Full” assessment, Kimi-K2.6 scored 54 points. This benchmark consists of over 2,500 doctoral-level questions across more than 100 academic fields. Moonshot AI reports that Claude Opus 4.6 scored 53 points and GPT-5.4 scored 52.1 points on the same test.

This release indicates that open-source AI competition is becoming increasingly fierce. While top-tier closed-source models still dominate the market, the rise of open-source models like Kimi-K2.6, which emphasize efficiency and task automation, is expanding options for businesses and developers.

TP AI Notice: This article is summarized based on TokenPost.ai’s language model. The main content may be omitted or may differ from actual facts.