Tencent’s new Hy4 model adds visual understanding and computer-use tools to its long-context system. The company says final model weights and system cards will follow in the coming days.
Tencent launched Hy4 on Nov. 4, 2026, adding visual understanding, long-context reasoning and computer-use capabilities that span command-line and graphical interfaces. The model is already accessible through selected products and API services, but Tencent says the final model weights and system cards will be released in the coming days.
The distinction matters. Developers can begin testing Hy4's capabilities through hosted services, while independent evaluation and deployment of the final model still depend on its promised release artifacts.
What is new in the final launch
Hy4 introduces visual understanding for the first time in the Hy line, according to Tencent’s launch announcement.
Tencent says the model can process detailed images, analyze complex charts and handle mathematical visual reasoning. The company also reports improvements in instruction following, long-context comprehension and long-horizon agentic tasks.
Computer use is another major addition. Tencent says Hy4 can interact with both command-line interfaces and graphical user interfaces, allowing it to work with software rather than only generate responses or code.
The company says specialized training included professional applications such as Blender and CAD, targeting workflows that require multiple steps and interactions with complex software.
For developers, this could expand the range of tasks handled by AI agents, particularly where existing applications lack convenient APIs. However, the announcement does not independently establish how reliably Hy4 completes those workflows.
API access is arriving before the final weights
Tencent says Hy4 is available through Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, Yuanbao and other company services. It is also making the model's API available through third-party platforms including OpenRouter, Hermes and OpenCode, while planning additional product integrations.
Vercel's AI Gateway update, published Nov. 3, confirms that Hy4 is available through its platform under the model identifier tencent/hy4.
Vercel lists reasoning, tool calling, structured outputs, image input, a 1 million token context window and computer-use support. It also says Hy4 improves on the preview in instruction following, long-context work and extended tasks.
These listings establish a path for developers to access the model. They do not independently validate Tencent's broader performance claims.
The August preview provides scale, not final-model proof
Tencent's Aug. 28 preview described Hy4 as a 770 billion parameter model with 49 billion active parameters and a context window exceeding 1 million tokens.
Those specifications apply to the preview. They should not be treated as a separate, independently audited characterization of the final release.
In an internal blind evaluation involving 163 experts and 203 engineering tasks, Tencent reported that Hy4 Preview scored 2.99 out of 4.
The company reported comparison scores of 2.92 for GLM-5.3 and 2.94 for Kimi K3.
Tencent also reported a 31.8% end-to-end throughput improvement against its baseline following inference-system optimizations carried out with assistance from the preview model. That result does not establish an independently verified speed improvement for the final Hy4 release.
The preview release identifies the blind evaluation as internal. Its results illustrate Tencent's reported competitive position, not a consensus from independent benchmarks.
Why the open-source distinction matters
Tencent describes Hy4 as joining the world's leading open-source AI models. The August preview was already released with downloadable weights under an Apache 2.0 license, as documented in its Hugging Face repository.
The final release has a different status. Tencent says its weights and system cards will follow in the coming days.
Until those materials appear, developers cannot independently deploy or inspect the final checkpoint using publicly released weights.
The system card should also provide important evidence about the model's assessed capabilities, safety limitations and release safeguards. The final weights and documentation will be needed to examine architecture details, licensing and the reproducibility of Tencent's claims.
Computer use also raises questions beyond model performance. A system that can operate graphical interfaces or execute commands may affect files, applications and workflows directly. That makes permissions, error recovery and oversight important considerations for deployment.
A new competitor in an established field
Hy4 enters a market where computer-use capabilities are already available from competing AI developers.
Anthropic introduced computer-use support for Claude 3.5 Sonnet in October 2024, allowing developers to automate screen interactions through its API. OpenAI followed with its Computer-Using Agent research preview in January 2025.
Those earlier releases demonstrate that computer use is not unique to Hy4. They do not establish which model performs better, particularly because computer-use evaluations depend on the tasks, tools and environments being tested.
Sources: Anthropic's computer-use announcement and OpenAI's Computer-Using Agent research.
Tencent's potential differentiation lies in combining long context, multimodal understanding, command-line execution and graphical computer use within a model it intends to release with accessible weights.
Whether that combination delivers a practical advantage remains an open question.
What settles it
The immediate evidence to watch is Tencent's release of the final Hy4 weights and system cards.
Those artifacts should help clarify the final architecture, license, safety limits and conditions under which its performance claims can be reproduced.
For developers, the significant question is not simply whether Hy4 can interpret images or operate interfaces. It is whether the model can complete complicated tasks reliably, at practical cost, with controls appropriate for real software environments.
Tencent has established product and API availability for Hy4. Independent validation of its final-model performance and safety remains the next test.