Tencent open-sources Hunyuan Hy4 preview in an open-source AI push

Tencent has released and open-sourced Hy4 preview, the latest addition to its Hunyuan AI family, according to IT Brief UK. The launch puts Tencent in the latest round of competition over open-source large language models, as companies seek to improve reasoning performance while limiting inference costs — a trade-off that has become more prominent as businesses weigh whether advanced AI can be deployed at scale without sharply higher computing bills.

On specifications, Hy4 preview uses a Mixture-of-Experts design with 770 billion total parameters and 49 billion active parameters. It supports a context window of more than one million tokens, allowing it to process long documents, software codebases and large datasets in a single prompt.

For distribution, Tencent is making the model available through several of its products, including WorkBuddy, CodeBuddy, Yuanbao and ima. Developers can also access it through an application programming interface on Tencent Cloud TokenHub and OpenRouter.

Hy4 preview was developed for software engineering, office tasks, financial analysis, scientific research and game development. For software development it was trained to improve long-context understanding, planning, debugging and validation while supporting front-end work; for office and analytical tasks, Tencent says it performs better in complex working environments and financial analysis, with tuning for data analysis and cross-document collaboration; in game development it can generate a playable prototype from a single natural-language request and work with game engines through multi-turn interactions; and in scientific research Tencent says it improved performance in AI research and development, molecular dynamics simulation, condensed-matter physics and fundamental mathematics.

Strikingly, the model also played a role in its own development. According to Tencent, Hy4 preview took part in automated optimisation of training methods, data strategies, evaluation frameworks and low-level operators by proposing approaches, running experiments and iterating on the results — an early recursive self-improvement loop. The model also analysed bottlenecks in its own inference system and went through several rounds of optimisation, including operator fusion and communication, raising end-to-end throughput by 31.8% from a baseline measure.

On evaluation, Tencent framed the release around practical use rather than benchmark scores alone, but it also published internal comparative data: in a blind evaluation involving 163 experts and 203 engineering tasks, Hy4 preview achieved an average score of 2.99 out of 4.00, ahead of GLM-5.3 at 2.92 and Kimi K3 at 2.94.

Pricing and access are another focus. Tencent set API pricing for Hy4 preview at $0.834 per million input tokens, $2.501 per million output tokens and $0.042 per million tokens for cache hits, and is offering free access on WorkBuddy and CodeBuddy for two weeks while extending free access to Hy3 on both platforms. The company says the preview release is intended to gather feedback before broader availability, following a preview-first approach that lets models be refined through real-world usage.

With inference cost now a defining constraint for enterprise-scale deployment, releasing an ultra-large MoE model as open source at aggressive API prices gives developers an early signal of the economics Tencent expects to offer as Chinese and US AI groups compete. The things to watch next are developer and enterprise feedback, the timing of the full Hy4 release, and whether the million-token context and recursive self-improvement claims hold up in real engineering and office workloads.