Tencent Hunyuan 3 Officially Released: 295B Params, Search Matches GPT-5.5, Hallucination Halved
Yao Shunyu's first major deliverable at Tencent: MoE architecture with 295B total params, search capability matching GPT-5.5, hallucination rate cut in half.
On July 7, Tencent officially released Hunyuan 3 (Hy3). This is the first major AI product launch since Yao Shunyu (former Tsinghua University Yao Class professor) joined Tencent. The model uses MoE (Mixture of Experts) architecture with 295B total parameters, 21B activated per token, plus a 3.8B MTP (Multi-Token Prediction) layer.
Architecture Details
- MoE structure: 192 experts with top-8 selection, effective parameters roughly half of GLM 5.2
- Attention: 80-layer GQA grouped attention, 8 KV heads out of 64
- Hidden dimension: 4096, intermediate layer 13312
- Context window: 256K
- Vocabulary: 120,832
- Precision: BF16
The architecture is identical to the previously released Hy3 preview. Tencent attributes the performance gains to "improved post-training data quality and diversity" and "expanded RL compute scale," rather than architectural changes.
Performance
Tencent claims Hy3's search capability matches GPT-5.5, with hallucination rate halved compared to the preview version. This makes Tencent the first Chinese AI company to reach GPT-5.5-level performance on search benchmarks.
Yao Shunyu's arrival is considered a key factor. His research at Tsinghua covered large model training optimization and RL alignment — directly aligned with Hy3's improvement path of "post-training data quality" and "RL compute scale."
Industry Impact
Hunyuan 3's release marks a new breakthrough for Chinese LLMs in MoE efficiency. With 295B total but only 21B activated parameters, inference costs are far lower than a dense model of equivalent performance. Against the backdrop of GLM 5.2 already triggering "AI margin collapse" discussions, Hunyuan 3 further compresses per-unit inference costs.
Meanwhile, search capability matching GPT-5.5 means Chinese models can now compete head-to-head with top international models in RAG and web search scenarios. The halved hallucination rate directly addresses the reliability issue most criticized in enterprise deployment.
本文由无人日报（Deskless Daily）编译员整理发布。
Top comments (0)