Tencent Open-Sources Hunyuan Hy4 preview with 770 Billion Parameters

Tencent has unveiled and open-sourced its new large language model, Hunyuan Hy4 preview. The model has 770 billion parameters, of which 49 billion are active, and supports a context window exceeding 1 million tokens. According to Tencent, Hy4 preview ranks among the top tier of open models and was built with real production tasks in mind — programming, office work and scientific research.

The model is already available both in open form and through Tencent's products — WorkBuddy, CodeBuddy, Yuanbao and ima — as well as via API through Tencent Cloud TokenHub and OpenRouter. On the first two platforms the model is offered free of charge for two weeks, and access to the previous version, Hy3, has been extended until September 30.

Betting on productivity

Hy4 preview was trained on data prepared together with Tencent's specialists in software development, gaming, finance and security. In programming, its creators say, the model does a better job of planning, debugging and reviewing across long contexts, as well as with frontend work. In office scenarios it is geared toward data analysis, financial analytics and working with documents, spreadsheets and presentations. In game development, Hy4 preview promises to assemble a playable prototype from a single text prompt, and in science it aims to help with tasks ranging from molecular dynamics to fundamental mathematics.

In an internal blind test by Tencent involving 163 experts and 203 engineering tasks, the model scored an average of 2.99 out of 4.00, edging out GLM-5.3 (2.92) and Kimi K3 (2.94).

A model that helps build itself

A curious detail: according to Tencent, Hy4 preview was the first to take part in optimizing its own development — it proposed approaches to training methods, data-handling strategies and low-level operators, ran experiments and iteratively improved the results. The company calls this an early loop of recursive self-improvement. In addition, the model independently analyzed the bottlenecks in its own inference system and, after several rounds of optimization, raised end-to-end throughput by 31.8% relative to the baseline configuration.

API pricing is $0.834 per million input tokens, $2.501 per million output tokens and $0.042 per million tokens on a cache hit. The next models in the Hy4 series, Tencent promises, will be released in the near future.

Source: Hacker News