Tencent has open-sourced Hy3, a 295B-parameter Mixture-of-Experts model featuring 21B active parameters, following feedback from over 50 product teams. The release includes both standard and FP8 quantized weights on Hugging Face, ModelScope, GitCode, and CNB.
- Hy3 significantly outperforms similar-size models and rivals flagship open-source models with 2-5x the parameters.
- Hallucination rates dropped from 12.5% to 5.4%, and commonsense error rates fell from 25.4% to 12.7% through fine-grained data cleaning.
- Multi-turn intent tracking issues decreased from 17.4% to 7.9%, with long-dialogue benchmark scores rising from 42.9% to 75.1%.
- Tool-call success rates and error recovery improved, maintaining accuracy variance within 4% across scaffolds like CodeBuddy and Cline.
- The model is available under the Apache License 2.0 for deployment via vLLM or SGLang with OpenAI-compatible APIs.
These improvements position Hy3 as a reliable, cost-effective option for productivity tasks such as coding, document processing, and financial analysis.