Tencent launches hosted Hy4 while final model weights remain pending
Tencent launched Hy4 through hosted services with visual understanding and computer use, but its final model weights and system cards were not available at launch.
Tencent launched Hy4 on November 4 as a hosted model with visual understanding, tool calling and computer-use capabilities. The company said the final model weights and system cards would follow “in the coming days,” leaving those downloadable materials out of the launch-day release.
At launch, users could access Hy4 through Tencent Cloud TokenHub and Tencent products including WorkBuddy, CodeBuddy, Yuanbao, WeCom AI Assistant and ima. Tencent also named OpenRouter, Hermes and OpenCode as third-party API channels. An OpenRouter listing was live when the release was reviewed, but access through Hermes and OpenCode was not separately verified.
Hy4 can accept visual inputs and combine visual understanding, reasoning and tool calls for detailed image processing, chart analysis and visual mathematics, according to Tencent. The company said it trained the model to use professional software including Blender and computer-aided design tools through command-line and graphical interfaces.
Compared with the earlier Hy4 preview, Tencent said it expanded pre-training and post-training to improve instruction following, long-context comprehension and long-horizon agentic work. Those improvement and performance statements remain company claims; the opened sources did not include an independent laboratory evaluation of the full model.
The Hy4 preview released on August 28 already had public weights. Tencent described the preview as having 770 billion total parameters, 49 billion active parameters and a context window exceeding one million tokens. Its November 4 announcement did not confirm those specifications for the final Hy4 model or state the release license.
Tencent also said a Hy4 Flash variant would arrive soon. It gave no firm publication date or repository address for Hy4 Flash, the final Hy4 weights or the system cards.
More news

DeepSeek launches V4.1 Flash with native multimodal support

Liquid AI releases LFM3 family for on-device multimodal AI

Moonshot launches Kimi K3.1 Pro, plans weights for November 4
