Tencent has released the full version of Hy4, its flagship language model, adding image understanding and computer use to the text-only preview that came out in August. The company says it reads fine detail in pictures and complex charts, and can work across command-line and graphical interfaces. Training included Blender, CAD and multi-application workflows. Access is through Tencent Cloud, OpenRouter and Tencent's own apps, including CodeBuddy, WorkBuddy, Yuanbao and ima.
The size is unchanged from the preview: 770 billion parameters in total, 49 billion active, and a context window of about a million tokens. OpenRouter lists text and image input with text output. Tencent reports gains of 19.1% in game development and 9.8% in game logic over the preview, without giving the baselines, and says internal search tests cut hallucinations from 15.76% to 5.45% and speculative errors from 23.1% to 5.6%.
Pricing is where the outlets differ. Unite.AI and TechNode give $0.834 per million input tokens, $2.501 for output and $0.042 for cached input, the same as the preview, while SiliconANGLE rounds them to about $0.90, $2.70 and $0.045. Tencent Cloud's TokenHub takes 50% off until 31 December, and WorkBuddy and CodeBuddy carry a limited free quota from 4 to 15 November.
Screenshot that SiliconANGLE ran with its Hy4 report
Image: SiliconANGLE
In a finance test run by Tencent, with 213 professionals and 579 tasks, Hy4 scored 3.04, ahead of DeepSeek-V4.1-Flash at 2.97 and Kimi K3.1 Pro and GLM-5.3 at 2.68 each. Tencent also admits a gap to the best proprietary models on research-level maths, long-horizon coding, self-improvement and hard visual questions. Tencent adds that its Hyra 2.0 harness, an autonomous research agent, optimised pre-training infrastructure for smaller models. The update targeted the preview's habit of over-reasoning and over-checking its own answers.
About Hy4
Hy4 is the large language model from Tencent's Hunyuan team, with 770 billion parameters in total and 49 billion active, and a context window above a million tokens. It is aimed at coding, office work, data analysis, game development and research.
Tencent open-sourced a preview on 28 August 2026, offered through WorkBuddy, CodeBuddy, Yuanbao, ima and an API on Tencent Cloud TokenHub and OpenRouter. In a blind test Tencent ran with 163 experts and 203 engineering tasks, the preview scored 2.99 out of 4, against 2.92 for GLM 5.3 and 2.94 for Kimi K3. Tencent also said the preview helped tune parts of its own training and inference systems, lifting end-to-end throughput by 31.8% over a baseline.
Nov 2026
Released
Tencent says the weights and a system card will follow within days, with vLLM and SGLang support, and that a cheaper, faster Hy4 Flash is planned. The preview had open weights when it launched on 28 August.
Sources · 3