model GLM-4.5
https://z.ai/blog/glm-4.5
The new agentic model GLM-4.5, developed by the Chinese company Z.ai (formerly Zhipu AI), significantly improves the capabilities of complex code generation presented in the previous version of GLM-4. The model can now create complex things such as interactive mini-games and physics simulations in HTML, SVG, Python, and other formats.

It has a 355b MoE size, with 12b active parameters. It demonstrates very good results in solving real engineering problems and creating interfaces. It demonstrates reliable mastery of multi-stage tool use and dynamic code generation – surpassing previous leaders among open models, Kimi K2 and Qwen3-Coder.

On the chat https://chat.z.ai/ there is a separate button for creating full-stack applications. The announcement shows a Pokémon catalog as an example. The chat is free; API token prices are cheaper compared to the average.

horizon-alpha horizon-beta
https://openrouter.ai/openrouter/horizon-beta
This week, an unknown provider https://openrouter.ai/provider/stealth released the Horizon model for testing, first the alpha, and now the beta version. Usage is free. The model generates interfaces well and rose to 4th place in the "Programming" ranking on OpenRouter.

Those who studied the tokenizer note that it resembles the one used by Qwen; it might be the future Qwen4.

#newllmmodel