Tencent releases full Hy4 model with vision and computer use, but the weights are not open yet

Tencent released the full version of its Hunyuan Hy4 model on November 4, succeeding the August 28 preview: a 770-billion-parameter mixture-of-experts system with about 49 billion active parameters and a roughly 1-million-token context window, adding vision understanding and computer-use capability for the first time in the Hunyuan line. It is available via Tencent Cloud TokenHub, OpenRouter and Tencent products, but the model weights are not yet downloadable; Tencent says it plans to open-source them, with no date given. (Cover photo is a file photo of Tencent headquarters in Shenzhen.)
Tencent released the full version of its Hunyuan Hy4 model on November 4, succeeding the preview released on August 28. According to Tencent's announcement and multiple tech outlets, the biggest changes are two firsts for the Hunyuan line: vision understanding, and the ability to operate a computer. Tencent says Hy4 ranks among the strongest open models for the first time — a company claim, not yet backed by independent benchmarks, and best read as such. (Cover photo is a file photo of Tencent headquarters in Shenzhen.)
Two new capabilities: seeing images, operating a computer
On vision, Hy4 can handle fine-grained image details, complex charts and maths problems presented as images. On computer use, it can work across command-line and graphical interfaces to complete tasks spanning applications. In Tencent's own benchmark table, Hy4 scored 78.9 on the OSWorld 2.1 computer-use benchmark, below Claude Opus 5.5 at 81.8, just under Kimi K3.1 Pro at 79.6, and above GPT-6.1 at 75.2 — again company-reported figures, not independently reproduced. Tencent also says its internal evaluations put Hy4's hallucination rate at 5.45%, down from 15.76% in the preview.
Scale and where it is available
Hy4 is a mixture-of-experts system with 770 billion total parameters and about 49 billion active per inference, with a context window of roughly 1 million tokens. From launch day it is available through Tencent Cloud TokenHub and OpenRouter, and inside Tencent products including Yuanbao, WorkBuddy, CodeBuddy, ima and the WeCom assistant Dayuan. Official pricing is 6 yuan per million input tokens and 18 yuan per million output tokens; WorkBuddy and CodeBuddy are free from November 4 to 15, and TokenHub is half price through December 31.
Not open yet: a hosted release for now
The catch is that the full version can only be reached through Tencent's hosted services and partners for now: Tencent says it plans to release the model weights and a fuller system card later, but gives no date, and as of the morning of November 4 its public model hub still listed only the August preview. Until the weights appear, developers cannot download, run or fine-tune the full Hy4 themselves. Tencent also promises a lighter, faster Hy4 Flash version soon, again with no specs or date.
N環球 
💬 Comments
No comments yet. Be the first!