Tencent has released Hy3, an open-source 295 billion parameter Mixture-of-Experts (MoE) model designed for complex reasoning, agentic workflows, and long-context tasks. The model activates only 21 billion parameters per token, maintaining efficiency while supporting a 256K context window. Hy3 demonstrates strong performance on various benchmarks, including coding, STEM, and reasoning tasks, and has undergone specific training to improve reliability in tool calling, reduce hallucinations, and enhance multi-turn intent tracking. The model is accessible via an OpenAI-compatible API and offers a `reasoning_effort` parameter to control its thinking process. AI
IMPACT Sets a new benchmark for open-source MoE models, potentially accelerating research and development in agentic workflows and long-context applications.
RANK_REASON Frontier-lab model release with system card and benchmark scores. [lever_c_demoted from frontier_release: ic=1 ai=1.0]
- Apache License 2.0
- GLM-5.1
- GPQA Diamond
- IMOAnswerBench
- Mixture-of-Experts (MoE)
- Multi-Token Prediction (MTP)
- OpenAI
- OpenRouter
- SGLang
- SWE-Bench Multilingual
- SWE-Bench Pro
- SWE-Bench Verified
- Tencent
- Terminal-Bench 2.1
- USAMO 2026
- vLLM
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
