AI Affairs, home

Wednesday 4 November 2026

Technology · Brief

Tencent releases 770-billion-parameter Hy4 with visual reasoning

The open-weight model powers an update to WorkBuddy, while Tencent’s own evaluations put it ahead of two rivals on the tasks tested.

Glass office towers and a curved building at Tencent’s Shenzhen headquarters campus.
Photo: Lhzss8, CC BY-SA 4.0, via Wikimedia Commons (cropped)

Tencent released Hy4, a 770-billion-parameter open-weight AI model that can process images and reason about visual material, on 4 November, the South China Morning Post reported. The company has used it to update WorkBuddy, its workplace AI assistant.

Tencent says Hy4 can analyse complex charts and work through mathematical problems presented visually. Its predecessor, Hy3, had 295 billion parameters. On Tencent’s internal evaluations of WorkBuddy tasks, Hy4 recorded an average score of 83.18 and a pass rate of 96.15 per cent.

Reading a chart at work could become part of a request to the assistant, rather than requiring its contents to be described in words first.

On benchmarks Tencent published, Hy4 outperformed Z.ai’s GLM-5.3 and DeepSeek-V4.1-Flash-1017 across the tasks tested. Tencent also acknowledged that the model trails leading closed-source systems on demanding mathematics, extended coding and advanced visual questions. During training, the company filtered material concerning exploit development, reverse engineering and penetration testing to restrict the model’s hacking capabilities.

Sources

Topics: Agents, Enterprise adoption, Foundation models, Open source