Shenzhen and San Francisco, November 11, 2026 – Tencent HY and Inferact, the company founded by the creators of vLLM, today announced a strategic collaboration on Hy4, Tencent’s newest flagship model. Inferact operates Hy4 in production, and the two teams are jointly building native vLLM support for Hy4 Preview, the open-weight release of the model.
Inferact Platform serves Hy4, providing the infrastructure and engineering required to serve live traffic under defined performance, accuracy, and reliability requirements. Tencent leads the development of Hy4 and Hy4 Preview, including Hy4 Preview’s native integration with open-source inference engines.
“Open models create the most value when developers can deploy them reliably. By combining Tencent’s model engineering with Inferact’s production-serving expertise, we are making Hy4 dependable at scale and giving developers a production-tested path to run Hy4 Preview,” said Dick Zhu, Head of Hy Inference Infra, Tencent.
Production serving for Hy4
Inferact already operates production serving for Hy4, supporting live traffic across long-context reasoning, tool use, multimodal inputs, and sustained multi-turn interactions.
Before carrying production traffic, the serving path is tested against Tencent’s API contract, model-quality requirements, and performance targets. Tencent retains control of the checkpoint, serving requirements, and validation standard, while Inferact is responsible for operating the deployment against those requirements.
Once the deployment is live, Inferact manages capacity, routing, caching, observability, and failure recovery while continuously tuning performance as demand changes. This allows the serving system to accommodate new checkpoints and evolving workloads without changing Tencent’s reference behavior.
The result is a managed production path for Hy4. Tencent can expand access to the model without building and operating every layer of the serving stack, and enterprises and developers get a dependable Hy4 endpoint without running the distributed infrastructure behind it.
“A model’s real performance is the experience it delivers over time,” said Simon Mo, Co-Founder and CEO of Inferact. “Our job is to preserve the behavior Tencent designed while making the system resilient as traffic, contexts, and applications change. That is what turns a frontier model into infrastructure teams can depend on.”
Open-source collaboration built around production
Hy4 Preview is a large mixture-of-experts model built for long-context reasoning, tool use, and agentic applications. Its public serving path includes support for a context window exceeding one million tokens, reasoning and tool-call parsers, speculative decoding, and distributed execution.
Tencent and Inferact co-lead Hy4 Preview’s native integration with vLLM. Tencent contributes the architecture, reference behavior, and model-specific requirements, while Inferact develops and hardens the engine support across model loading, parallel execution, cache management, parsing, and performance-critical runtime paths.
The collaboration extends beyond enabling the model to run. The teams are developing validation coverage, deployment configurations, performance guidance, and reproducible recipes so developers can move from downloading the Hy4 Preview checkpoint to operating a complete serving system.
Findings from serving Hy4 in production that have broader value can be incorporated into Hy4 Preview’s public implementation or upstreamed into vLLM. Likewise, improvements from the open-source community can strengthen future production configurations, keeping the open and operated serving paths aligned as both evolve.
Why this matters
Open-weight releases have made frontier models more accessible, but operating them at scale remains difficult. A public checkpoint may be available on launch day, while the production path—engine support, deployment architecture, performance tuning, and validation—remains unfinished.
Tencent and Inferact are closing that gap by developing Hy4’s production deployment and Hy4 Preview’s open engine integration as parts of the same lifecycle. The open serving path is informed by the model developer’s requirements and by the conditions that emerge when Hy4 runs at scale.
For Hy4 users, that means faster access to a dependable endpoint. For developers operating Hy4 Preview themselves, it means starting from a serving path shaped by real production experience. For the open-source ecosystem, it means that production requirements help shape the infrastructure everyone can use.
What to expect next
Tencent and Inferact will continue expanding Hy4’s production deployment as traffic grows, while validating new checkpoints and serving configurations against Tencent’s requirements.
The teams will also continue improving performance for long-context inference, tool use, caching, and distributed serving across hardware platforms. Production will remain part of the development process: findings from real workloads will inform deployment guidance for Hy4 Preview and future open releases, and improvements with broader value can flow back into vLLM and its supporting ecosystem.
As Hy4 and Hy4 Preview evolve in capability and scale, the collaboration will keep Tencent’s open releases and production deployment moving forward together, so that new model capabilities become usable in production sooner.
About Inferact
Inferact is building the production inference infrastructure for AI. Founded by the creators and core maintainers of vLLM, the world’s most widely adopted open-source LLM inference engine, Inferact makes it faster, cheaper, and easier to serve frontier AI models at scale, across models and hardware, on infrastructure companies control. Learn more at inferact.ai.
Media: contact@inferact.ai
