Tencent’s Hy4 strengthens open-model competition but trails closed leaders

Hy4 expands Tencent’s open-model proposition

Tencent announced the release of Hy4 on November 4, describing the model as among the strongest open-source systems while acknowledging that it remains behind leading closed models. The release adds visual understanding and computer use, alongside improvements in instruction following, long-context comprehension, long-running agent tasks, and deep research.

Hy4 went live in Tencent’s Yuanbao, WorkBuddy, CodeBuddy, ima, and WeCom assistant Dayuan products, as well as through Tencent Cloud’s TokenHub. Developers outside China can also access it through OpenRouter. The full model was released as a hosted service, but Tencent said it only plans to publish the model weights and system card in the coming days.

That distinction matters for the open-model market: Hy4 is commercially available through hosted channels, while its promised open-weight distribution has not yet occurred. Tencent also has not specified the license for the full model’s weights.

Computer use is the central capability jump

The most consequential addition is Hy4’s ability to interact with command-line tools and graphical user interfaces. Tencent said the model was trained on professional software, including Blender and CAD tools, as well as workflows spanning several applications. The capability extends the model’s intended role from answering questions and analyzing content toward planning and executing tasks within software environments.

Tencent’s internal evaluation illustrates both the promise and the limits. Hy4 scored 78.9 on OSWorld 2.1, compared with 81.8 for Claude Opus 5.5 and 75.2 for GPT-6.1 Sol. Its score was lower on Terminal Bench 4.0, at 44.4 versus 66.4 for Claude Opus 5.5 and 55.5 for GPT-6.1 Sol. These are Tencent-reported results, and several rely on tests built by the company.

Tencent also reported that Hy4’s hallucination rate fell from 15.76% for its preview to 5.45% in internal search evaluations. That suggests progress against the earlier release, but it is a company-controlled measure rather than independent validation.

Tencent’s product integration is the commercial advantage

Hy4’s immediate route to adoption runs through Tencent’s existing applications rather than a new standalone distribution channel. WorkBuddy evaluations reported by Tencent produced an average performance score of 83.18, a task pass rate of 96.15%, and first place in nine of 12 mainland Chinese office scenarios. Tencent also integrated Hy4 with WeCom workflow data, including chats, emails, documents, calendars, and meetings, while a forthcoming WorkBuddy “Projects” feature is expected to add team-collaboration capabilities.

The model also targets financial analysis. In a Tencent-run blind evaluation involving 213 financial experts and 579 real tasks, Hy4 averaged 3.04 out of 4, ahead of DeepSeek-V4.1-Flash-1017 at 2.97, while Kimi K3.1 Pro and GLM-5.3 each scored 2.68. These results support a plausible productivity use case, but they do not establish external customer demand or independent superiority.

Pricing may support early adoption

Tencent retained the preview’s TokenHub prices: 6 yuan, about $0.85, per million input tokens; 18 yuan, about $2.50, per million output tokens; and 0.3 yuan for cached input. Calls are half price until December 31. OpenRouter listed Hy4 at $0.42 per million input tokens and $1.25 per million output tokens, while Tencent Cloud was its only listed provider.

The reviewed materials show relatively low hosted token prices for Hy4, which could reduce friction for experimentation, although they do not establish how its pricing compares with proprietary alternatives. Its practical advantage is the combination of relatively low token prices, a million-token context window, and immediate availability inside Tencent’s productivity ecosystem. The reviewed reporting does not establish customer volumes, paid adoption, or realized revenue from those access channels.

Open-weight leadership remains unproven

Tencent’s claim that Hy4 has entered the top tier of open-source models is supported by its own and affiliated evaluations, but the evidence is not conclusive. It leads DeepSeek and Z.ai models on most of Tencent’s tests, while Moonshot’s Kimi K3.1 Pro outperforms it in several reported comparisons. Hy4 also remains materially behind Claude and GPT on the hardest internal tests involving research mathematics, long software projects, model development, and visual questions.

The next observable milestones are the promised publication of Hy4’s weights and comprehensive system card, including fuller safety results, and the release of a lower-cost “Hy4 flash” model that Tencent says is due soon. Until those steps occur, Hy4 is a released hosted model with a planned—not completed—open-weight distribution.

Industry and product implications