Tencent has released the Hy4 foundation model, saying it reaches comparable performance levels to the latest releases from Z.ai, Moonshot AI and DeepSeek in tests including programming, agents and browser use.

Hy4 introduces visual understanding for the first time, enabling it to process images, analyse complex charts and conduct visual mathematical reasoning, the Shenzhen-based company said.

The model is the first to turn Tencent’s AI range into “one of the strongest open-weight models available today”, the company said.

Tencent Seafront Towers in Shenzhen, China.

Competitive improvements

It acknowledged, however, that the model still trails leading proprietary releases from companies such as Anthropic and OpenAI on benchmarks including research-level mathematics, long-horizon coding, model self-improvement and advanced visual question answering.

The Hy4 model series includes 770 billion parameters, up 2.6 times from the 295 billion parameters of the Hy3 range.

The Hy5 range is planned to continue scaling at a comparable rate, while a lightweight Hy4 Flash model will cut task costs, Tencent said.

Reduced hacking capabilities

Tencent said Hy4 outperformed Z.ai’s GLM-5.3 and DeepSeek-V4.1-Flash-1017 across all tested tasks, and was comparable to Moonshot’s Kimi K3.1 Pro, according to benchmarks published with the release.

The company said it restricted the model’s offensive cyber-security capabilities during testing, a step also taken by Moonshot for Kimi K3.1 Pro.