Hy3 Preview Released

Tencent Hunyuan released the Hy3 preview and Hy3 preview-Base weights on 2026-04-23 on Hugging Face, ModelScope, and GitCode.

The model uses a 295B MoE architecture with 21B active parameters and 3.8B MTP layer parameters. It features 80 layers, 64 heads (GQA, 8 KV heads), hidden size 4096, intermediate size 13312, vocab 120,832, context length 256K, 192 experts (top-8 activated), and BF16 support.

• STEM / reasoning: The company stated it showed strong reasoning performance on FrontierScience-Olympiad, IMOAnswerBench, the Tsinghua Qiuzhen College Math PhD qualifying exam (Spring '26), and CHSBO 2025.
• Context learning / instruction following: The company reported improvements on the in-house scenario-based CL-bench and CL-bench-Life.
• Code / agent: The company disclosed competitive results on SWE-bench Verified, Terminal-Bench 2.0, BrowseComp, WideSearch, ClawEval, WildClawBench, as well as internal evaluations Hy-Backend, Hy-Vibe Bench, and Hy-SWE Max.

For the pre-trained comparison, it was benchmarked against Kimi-K2 BASE, DeepSeek-V3 BASE, and GLM-4.5 BASE, with representative figures given as MMLU 87.42, MMLU-Pro 65.76, ARC-Challenge 95.99, LiveCodeBench-v6 34.86, GSM8K 95.37, MATH 76.28, and MMMLU 80.15.

The deployment documentation includes vLLM and SGLang examples, instructions for calling the OpenAI-compatible API, 8 GPU serving, and a recommendation of H20-3e-class memory or higher. The training pipeline supports both full fine-tuning and LoRA, and includes DeepSpeed ZeRO and LLaMA-Factory integration guides.