Chinese tech giant Tencent has released its latest open-source AI model, Hy4, which it says ranks among the world’s best while maintaining limited cyber and biological offensive capabilities.

Hy4 scores on par with or above other leading open-source models such as Moonshot AI’s Kimi K3.1 Pro, Z.ai’s GLM 5.3 and DeepSeek-V4.1-Flash-1017 on a range of industry-standard benchmarks, according to Tencent’s data.

However, the company has also deliberately limited Hy4’s training dataset, omitting materials related to offensive cyber exploitation and offering weaker capabilities for open-ended experimental biological design than frontier models. This decision mirrors that of fellow Chinese developer Moonshot when releasing Kimi K3.1 Pro in October.

Governments, regulators and major AI developers have all expressed concerns around the offensive potential of advanced models in recent months, spurred by a series of incidents by leading US models during testing.

Outside of these areas, however, Tencent said Hy4 is extremely capable.

It is the company’s first open-weight model to offer visual reasoning and computer-use capabilities. This allows it to perform tasks such as fine-grained image processing and complex chart analysis, as well as interacting with command-line interfaces and graphical user interfaces.

Outside of coding capabilities, the model is also adept at scientific research, Tencent said, ranking above other open-weight models on an internal TenureTrail benchmark.

Tencent plans to release the model’s weights and system cards “in the coming days”, it said. For those running the model through Tencent’s API, its pricing remains the same as the previous generation at around 90 cents per million input tokens, $2.70 per million output tokens and 4.5 cents per million cached input tokens.

---