Tencent releases Hy4 with vision and computer use, but has not published the weights yet

What is new in Hy4

On Nov. 4, Tencent announced Hy4, its new flagship model. The company says the Hy line gains visual understanding for the first time: fine-grained image processing, complex chart analysis and visual mathematical reasoning. Computer use is also new; the model can work with both command-line and graphical interfaces, and was specially trained on professional software such as Blender and CAD tools. Tencent also reports gains in instruction following, long-context comprehension and long-horizon agentic tasks. The Hugging Face model card for August’s Hy4 preview puts the series at 770 billion total parameters with about 49 billion active per token, a mixture-of-experts design with a 1-million-token context. The South China Morning Post notes that is 2.6 times the figure for its predecessor, Hy3. The Hy4 announcement itself gives no parameter count.

The company’s own measurements

Every number in the 12-panel benchmark chart in the announcement is Tencent’s own measurement. On the computer-use test OSWorld 2.1, Hy4 scored 78.9, against 79.6 for Kimi K3.1 Pro and 81.8 for Claude Opus 5.5. On Terminal Bench 4.0, Hy4 reached 44.4 while Opus 5.5 scored 66.4 and Kimi 66.8. On ProgramBench the order is 30.0 for Hy4, 46.0 for GPT 6.1 Sol and 56.5 for Opus 5.5. Tencent acknowledges that it trails the world’s leading closed systems on harder internal benchmarks such as research-level mathematics, long-horizon coding, model self-improvement and advanced visual question answering, and says it will keep scaling pre-training and post-training to narrow the gap. The product figures are company measurements too: an average score of 83.18 and a 96.15% pass rate on WorkBuddy evaluations, and a hallucination rate down from 12.95% in the preview to 7.2%.

Weights, pricing and access

The announcement says model weights and system cards will be published “in the coming days.” As of Nov. 7 there is no Hy4 repository on the tencent account at Hugging Face; the only open version is August’s preview, under Apache 2.0. Tencent kept pricing level with the preview: about $0.90 per million input tokens, $2.70 per million output tokens and $0.045 per million cached input tokens. The model was added to OpenRouter on Nov. 4, where the context is listed at 1,048,576 tokens, input covers text and images, and the listed price is $0.417 per million input tokens and $1.25 per million output tokens. From Nov. 4 to 15 Tencent is making WorkBuddy and CodeBuddy free, and it is offering 50% off on Tencent Cloud TokenHub through Dec. 31.