On November 4, 2026, Tencent released Hy4 — its first flagship that, in addition to text, understands images and can control a computer: CLI terminal, graphical interfaces, mouse, and windows, with training including professional programs like Blender and CAD. The model's context reaches 1 million tokens with a maximum output of 65,536 tokens. Hy4 is already available via API on Tencent Cloud and OpenRouter with a temporary 50% discount on tokens; open weights are promised in the coming days, followed by a lightweight Hy4 Flash. Tencent's own benchmarks paint a mixed picture: Hy4 outperforms GPT-6.1 Sol on long agent tasks but loses to the same model in the terminal.

image
image
image

What happened

On November 4, 2026, Tencent released Hy4 — a new flagship multimodal LLM. For the first time at the flagship level, the company added image understanding and computer control: the model works with CLI terminal, graphical interfaces, mouse, and windows, and training included data from professional programs like Blender and CAD. Context size is 1 million tokens, maximum output is 65,536 tokens. The model is already available via API on Tencent Cloud and on OpenRouter under the identifier tencent/hy4: input costs $0.834, output — $2.501 per 1 million tokens, with a current 50% discount. Exact parameters of the new version are not disclosed; open weights are promised later — "in the coming days" — while a lightweight version, Hy4 Flash, is being prepared in parallel.

Context

The previous Hy4 preview version had 770 billion total and 49 billion active parameters in a MoE architecture, so the new flagship likely belongs to the same class of sparse models, although exact parameters are not disclosed and it is impossible to judge the efficiency of scaling. Benchmarks provide their own measurements from Tencent: on E-Bench 2.0, Hy4 scores 70 points compared to 65.3 for GPT-6.1 Sol; on Terminal-Bench 4.0 — 44.4 compared to 55.5 for the same model; on APEX-Agents — 75.2%; on OSWorld 2.1 — 78.9%. These figures are only partially consistent: the model outperforms GPT-6.1 Sol on long agent tasks but lags behind it in the terminal. Notably, Tencent publishes the losing metric — this reduces the risk of total selection of favorable results, although it does not replace independent verification. From the combination of these numbers, the main analytical conclusion follows: GUI agency and reliability in the console are different skills, and the general term "computer use" masks them. Finally, Tencent presents Hy4 in the ranks of the world's leading open-source models, which sets a high level of expectations for the entire lineup.

Why this matters for the industry

For the industry, Hy4 is a new building block: the combination of vision, computer use, and 1 million token context is available via API at half price, which reduces the cost of prototyping clicking agents and vertical assistants for Blender and CAD. This strengthens competition in the segment of agent and long-horizon tasks, where Tencent claims leadership on APEX-Agents and near top-level results on OSWorld 2.1. The gap between the ability to control a computer and reliable work in the terminal is a specific benchmark for choosing a model for agent pipelines: terminal tasks are best directed to other models. Promised open weights will open up self-hosting and fine-tuning, including work with sensitive data, and the release of Hy4 Flash will change the economics of routing — typical requests can be transferred to the cheaper version. If the trajectory is maintained, multimodality and computer use will become a mandatory set for flagships, and the division of "heavy flagship plus lightweight Flash version" will become the norm for inference economics on a two-year horizon.

Why this matters for users

You can try Hy4 right today: on OpenRouter under the identifier tencent/hy4 or via Tencent Cloud API — while the 50% discount is active, experiments cost approximately half the base price. On this model, a pilot computer-use product is being built on the same day: an agent cycle with screenshots and mouse actions for GUI scenarios, working with long documents on a 1 million token context, assistants for Blender and CAD. It is rational to set up an eval harness on your own cases and run it before the discount is canceled — then the decision on implementation will be based on your own measurements, not Tencent's figures. For terminal and CLI scenarios, Hy4 is not yet the best choice: the declared metric on Terminal-Bench is significantly lower than the level of GPT-6.1 Sol. Further, it is worth tracking the release of open weights, which will allow self-hosting and fine-tuning, and the lightweight Hy4 Flash, which will reduce the cost of mass agent products.

What is currently unknown / limitations

All key figures are Tencent's own measurements without disclosed methodologies; there is no independent verification in the sources, so until contractual obligations, results should be perceived as declared, not confirmed. Exact parameters of Hy4 are not disclosed, and conclusions about scaling compared to Hy4 preview cannot be made. The claim to be among the leading open-source models will remain unverified until open weights are actually released — until then, open-source status is formal. The 50% discount is temporary, the price after its cancellation is unknown, so it is risky to build the economics of long-term products on the current tariff. Finally, the model's weakness in the terminal is confirmed only by Tencent's own benchmarks and requires re-verification on your own scenarios.

Sources

Author

Look at AI, editorial team