By Paul Derrington, Jr.
CEO | Chief Engineer | Research Engineer | Editor, D-CSIL
3 min read
Key Takeaways
Tencent released the full version of Hy4 on November 4, adding vision and computer use to the model it previewed in August, and it is live now through Tencent Cloud, OpenRouter, and Tencent's own apps at unchanged prices. Tencent says the model ranks among the strongest open-weight models for the first time, and reports lower hallucination rates in its own search tests, though its two English announcements give different figures. The open weights are not out yet: the August preview shipped with its weights on day one, and Tencent says it plans to publish the full release's weights.
The preview: text only, open from day one
Tencent open-sourced the Hy4 preview in late August as a text-only model: 770 billion total parameters with 49 billion active in a mixture-of-experts design, and a 1-million-token context window. It was built for coding and office work, and Tencent itself flagged excessive reasoning and over-verification as known issues in that early version.
The preview's weights went up on Hugging Face the same day under the Apache 2.0 license.
The full release: vision, computer use, and a finance test
According to Tencent, the November 4 release adds vision and computer use: Hy4 can now read images and charts, and it can operate command-line and graphical interfaces, including 3D and CAD tools. Tencent says it trained the model on professional software such as Blender and CAD, and adapted it for game development in Unreal Engine, Unity, and Godot.
Tencent also ran a finance evaluation with 213 finance professionals, including 153 experts from financial institutions, across 579 real-world tasks. Tencent says Hy4 ranked first among open-weight models with a score of 3.04, ahead of DeepSeek-V4.1-Flash-1017 at 2.97 and Kimi K3.1 Pro and GLM-5.3 at 2.68 each. On hallucinations, Tencent's two English announcements disagree: its Hy research page says internal evaluations cut the rate from 15.76% on the preview to 5.45%, while its corporate press release gives 12.95% to 7.2%. Both say errors from unsupported assumptions fell from 23.1% to 5.6%.
Same price, same wait for the weights
Tencent says pricing is unchanged from the preview. Its Hy research page lists $0.834 per million input tokens and $2.501 per million output tokens, with cached input at $0.042; its corporate release gives approximate figures of $0.90, $2.70, and $0.045. Tencent Cloud's TokenHub is running a 50% discount through December 31, 2026 — OpenRouter's live listing shows the discounted $0.417 and $1.251 — and WorkBuddy and CodeBuddy are free to use from November 4 to 15. The model is also live inside Yuanbao, ima, and WeCom's AI assistant.
The weights have not yet been released. Tencent's corporate release says it plans to release the model weights and a comprehensive system card “in the coming days,” and its research page says it has partnered with the vLLM and SGLang communities for support. No date was given. As of November 4, Tencent's Hugging Face page still listed only the August preview, and OpenRouter shows Tencent Cloud as the model's only provider. Tencent says a more cost-efficient Hy4 Flash is “scheduled for release soon.”
On safety, Tencent says cybersecurity training focused on defensive proof-of-concept work and vulnerability patching. Its two announcements describe the rest differently: the research page says exploit development, reverse engineering, and penetration-testing content was filtered out, while the corporate release says the model incorporates reverse engineering and penetration testing “for analysis purposes” and excludes offensive exploitation material. Tencent says its Xuanwu Lab and Keen Security Lab red-teamed the model and found no elevated risk in the offensive tasks tested, and that it scored 30.4 out of 100 on Hy-BioExperimentBench. Reuters reported Tencent's description of the limits and its plan to release the weights with a safety report.