Tencent's AI news account on X says Hy4 is out of preview, adding native vision and computer use while keeping the preview's API prices. Tencent describes it as the first Hy model with eyes, and says open weights and a smaller Hy4 Flash model are coming.

Image credit: @TencentAI_News on X
Hy4 is a 770B-parameter mixture-of-experts model, according to ChainCatcher's report. The earlier Hy4 preview was text-only.
What changed from the preview
Vision: Hy4 adds visual understanding, according to ChainCatcher's second report.
Computer use: the model combines command-line and graphical interface control, which ChainCatcher says is a first for Hy models. Tencent's release page says it works across both interfaces.
Training and instruction following: Tencent says it expanded pre-training and post-training data and scale, and improved instruction following and long-text understanding.
Hallucinations: Tencent's internal evaluations put the hallucination rate at 5.45%, down from 15.76% on the preview, as FourWeekMBA reports.
Price and access
Hy4 keeps the preview's API pricing: 6 yuan per million input tokens, 18 yuan per million output tokens and 0.3 yuan per million cache-hit tokens, according to ChainCatcher. A third-party catalogue listing shows the same model at $0.834 per million input tokens, $2.501 per million output tokens and $0.042 for cached input. That listing is not Tencent's price sheet.
According to FourWeekMBA, Hy4 is offered through Tencent Cloud and OpenRouter APIs. It is also built into CodeBuddy, WorkBuddy, Yuanbao and ima.
Benchmarks
Tencent says Hy4 makes major performance gains over the preview at the same API price, and its release page calls the model one of the strongest open-source models. The benchmark figures below come from a third-party summary, not from Tencent. On Apiyi's comparison, Hy4's largest lead over DeepSeek V4.1 Flash is computer use, with OSWorld at 78.9 against 68.4.
Open weights and Hy4 Flash
Tencent says the Hy4 weights are coming soon, with a system card to follow when they are released, and that it has partnered with the vLLM and SGLang communities for the release. The preview's weights shipped on August 28 under Apache 2.0, as its GitHub repository shows. Tencent says Hy4 Flash will follow in the near future.
For API users, the same prices now buy vision and computer use. If you want to self-host, wait for the weights and the system card before planning around them.





0 comments
No approved comments yet. You can start the conversation.
Leave a comment