Key facts
- Tencent released three Hy4 checkpoints on Hugging Face, ModelScope and GitCode, including an FP8-quantized version and a base model.
- The system card reports CyberGym vulnerability reproduction rising to 84.87% from 74.05% for Hy4 Preview.
- Tencent's card reports a 35.9% over-refusal rate on benign agent tasks with reasoning off, against 13.5% for the preview.
- Tencent's launch post names its own Xuanwu, Zhuque and Keen labs alongside outside testers ALICE and Fangcun AI.
- Tencent says cyber training data was restricted to defensive material, but general coding let the model partially reconstruct offensive capabilities.
Tencent Holdings released the open weights for Hy4, its flagship large language model, on November 20, 2026, sixteen days after the model first launched through Tencent's apps and API on November 4. The release came with an 82-page system card covering safety, security and alignment evaluations, according to a blog post by orcarouter.ai.
The card reports gains in defensive cybersecurity. On CyberGym, vulnerability reproduction rose to 84.87% from 74.05% for the earlier Hy4 Preview, and on SEC-bench Pro it rose to 64.24 from 40.70, per orcarouter.ai. Tencent said cyber training data was restricted to defensive material, with exploitation and penetration testing stripped out, but that general coding ability let the model partially reconstruct offensive capabilities.
The same document reports a 35.9% over-refusal rate on benign agent tasks when reasoning is switched off, against 13.5% for the preview. Both figures are Tencent's own and have not been reproduced, orcarouter.ai said.
Tencent's launch post names its own Xuanwu, Zhuque and Keen labs alongside outside testers ALICE and Fangcun AI, orcarouter.ai reported, adding that the three Tencent units are on the company's payroll. Artificial Analysis, Collinear and Scale AI supplied external benchmarks.
According to themodelatlas.com, Tencent open-sourced three checkpoints on Hugging Face, ModelScope and GitCode: the instruction-tuned Hy4, an FP8-quantized version and a pre-trained base model. The model card lists 770 billion total parameters with 49 billion active per token, a 1-million-token context window and multimodal input covering text, image and video.
