Overview
- Tencent fully launched Hy4 on Nov. 4, 2026, describing it as a 770‑billion‑parameter mixture‑of‑experts model with a roughly 1,048,576‑token context window designed for long‑horizon, agentic workflows.
- Hy4 adds native vision and computer‑use abilities that let the model interpret images and operate desktop and command‑line software such as Blender and CAD to automate multi-step tasks.
- Tencent’s internal tests show Hy4 cut its hallucination rate from 12.95% to 7.2% and reduced errors from unsupported assumptions from 23.1% to 5.6%, and the company reported a top score of 3.04 on a 579‑task finance evaluation with 213 professionals.
- The company published commercial pricing and said Hy4 will be available through Tencent Cloud and product integrations, will ship a smaller Hy4 Flash for lower-cost use, and will release model weights and system cards in the coming days.
- Independent benchmarks place Hy4 behind the leading closed‑source frontier models and experts warn that GUI-based agent actions remain brittle, release of weights raises dual‑use and safety questions, and restricted access to advanced chips could shape Tencent’s deployment options.