Tencent and Inferact Partner on Hy4 Serving and vLLM Support

Tencent and Inferact, the company founded by the creators of vLLM, are collaborating on production serving for Hy4, Tencent's newest flagship model, and native vLLM support for Hy4 Preview, the companies announced on Inferact's website. Inferact already serves live Hy4 traffic for long context reasoning, tool use, multimodal inputs and conversations with multiple turns.

Tencent controls the model checkpoint, serving requirements and validation standards. Inferact manages capacity, routing, caching, monitoring and failure recovery while testing the service against Tencent's API contract, performance targets and model quality requirements.

The companies co lead native vLLM integration for Hy4 Preview, an open weight mixture of experts model. Tencent contributes the architecture and reference behavior, while Inferact develops the engine support for model loading, parallel execution, cache management and parsing. Hy4 Preview's public serving path supports a context window exceeding one million tokens, speculative decoding and distributed execution.

The teams are also developing validation coverage, deployment configurations, performance guidance and reproducible serving recipes. Findings from production use can be incorporated into Hy4 Preview's public implementation or upstreamed into vLLM.