Tencent Launches Hy4, a 770B Parameter Model With Vision, Promising Open Weights Later
On November 4, Tencent launched Hy4, the final release of its large mixture-of-experts model, introducing vision and computer-use capabilities for the first time. It is available via API starting today, while open weights remain merely promised.
Tencent launched Hy4 on November 4, 2026, the final version of the large language model in its Hy series (formerly Hunyuan). Compared to Hy4 preview, released in late August, the Shenzhen-based group states that it significantly increased pre-training and post-training, adding two previously missing capabilities: vision and computer use. The model can be used today via API; its open weights, however, are merely promised.
What is available today
Hy4 is accessible on Tencent Cloud and OpenRouter, and integrated into the group’s products: the code assistant CodeBuddy, the productivity tool WorkBuddy, the consumer assistant Yuanbao, and ima. Tencent lists pricing identical to that of the preview.
| Item | According to Tencent |
|---|---|
| Base | continuation of Hy4 preview: mixture-of-experts, 770 billion parameters, 49 billion active |
| Context | 1 million tokens (preview model card) |
| New features | vision, computer use (CLI and graphical interfaces) |
| Price per million tokens | $0.834 input, $2.501 output, $0.042 cached |
| Weights | promised, no date; preview weights are licensed under Apache 2.0 |
The model’s OpenRouter page displayed pricing this morning that was half the rate announced by Tencent; Tencent has not commented on this discrepancy. Detailed architecture has only been published for the preview version: 78 layers, 256 experts per layer with 8 activated per token, and sparse attention inspired by DeepSeek and GLM.
Progress largely measured internally
Tencent emphasizes practical applications over a single benchmark scorecard. The model was trained on specialized software such as Blender and CAD tools, tailored to the Unreal, Unity, and Godot game engines (+19.1% on the GameDev benchmark compared to the preview), and evaluated in finance by 213 professionals across 579 tasks, where it reportedly ranks first among open models ahead of DeepSeek V4.1 Flash, Kimi K3.1 Pro, and GLM-5.3. In information retrieval, its hallucination rate reportedly dropped from 15.76% to 5.45%. These figures come from Tencent or partner clients and have not been independently reproduced.
The only external signal so far comes from the Arena: according to its official account, Hy4 debuts at 14th place on its agent leaderboard, which would place it tied with Kimi K3.1 Pro at the top of open models once its weights are released. This ranking relies on user votes and shifts quickly.
Tencent itself acknowledges a “notable gap” with leading proprietary models in research-level mathematics, long-horizon coding tasks, and advanced visual reasoning. On the safety side, the group claims it removed offensive cybersecurity data from training and measured biological risk below its alert threshold; it promises a comprehensive system card upon the release of the weights. A Hy4 flash version, faster and cheaper, was announced without a date.
Our take
Hy4 confirms that Chinese tech giants are now battling for the top spot in open-weight models, following DeepSeek, Moonshot (Kimi), and Zhipu (GLM). However, the announcement must be read for what it is: a commercial launch via API, with open weights merely promised. Until those weights are actually published, Hy4 is not an open model, and comparisons to Kimi K3.1 Pro or GLM-5.3 remain Tencent’s own claims. Precedent works in its favor, given that the preview version was indeed released under Apache 2.0. For developers, the immediate appeal lies in the pricing—substantially lower than that of top American models—and the arrival of vision and computer use, which had been missing from the lineup.
Sources
- Tencent Hy — Introducing Hy4 (November 4, 2026) Primary source
- Tencent Hy on X — Hy4 announcement (November 4, 2026) Primary source
- Hugging Face — tencent/Hy4-preview model card Primary source — Architecture and license of the preliminary version
- OpenRouter — tencent/hy4 model card Primary source
- Arena on X — Hy4 in Agent Arena (November 4, 2026) Secondary source
- IT Home — 腾讯混元 Hy4 正式版发布 (November 4, 2026) Secondary source
- South China Morning Post — Tencent unveils Hy4 (November 4, 2026) Secondary source
A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.
