Tencent open-sources Hy3, 295B MoE model with 21B active
Tencent open-sourced Hy3, a 295B-parameter Mixture-of-Experts model that runs with 21B active parameters, scores 74.4% on SWE-bench Verified and is available on GitHub, Hugging Face and Tencent Cloud.
Tencent released Hy3 on Thursday as an open-source preview. The model contains 295 billion parameters in total and operates with about 21 billion active parameters per query using a Mixture-of-Experts (MoE) architecture. MoE routes each request to a small set of expert subnetworks rather than running the whole model, reducing runtime compute requirements.
Development began in late January 2026 and the preview was published in under three months after an infrastructure rebuild led by chief AI scientist Yao Shunyu. Tencent posted the model and weights on GitHub, Hugging Face and ModelScope, and it also offers commercial API access through Tencent Cloud.
Benchmark results released with the model show performance differences from the company’s previous generation. On SWE-bench Verified, which tests whether a model can fix real GitHub bugs, Hy3 scored 74.4%, up from 53.0% for Hy2. On Terminal-Bench 2.0 Hy3 scored 54.4% versus 23.2% for Hy2. On BrowseComp Hy3 reached 67.1% (Hy2 28.7%), and WideSearch measured 70.2%. On Tsinghua University’s math PhD qualifying exam (Spring 2026) Hy3 averaged 88.4 across three runs; on CHSBO 2025 it scored 87.8.
Tencent says Hy3 supports context windows up to 256,000 tokens. The company has deployed the model across its products, including Yuanbao, QQ and Tencent Docs, and made it available on the Openclaw agent platform for developers building multi-step, tool-using agents.
Internal deployments on CodeBuddy and WorkBuddy showed a 54% reduction in first-token latency and a 47% decrease in end-to-end generation time. Internal testing ran agent workflows up to 495 steps. Tencent lists API pricing at about $0.18 per million input tokens and $0.59 per million output tokens, with personal token plans starting near $4.10 per month.
Tencent described Hy3 as an effort to balance model capability, evaluation and cost-efficiency. The company noted that its prior flagship, Hy2, had more than 400 billion parameters and that Hy3 reduces total parameters while preserving capacity through expert routing and updated training techniques.
The material on GNcrypto is intended solely for informational use and must not be regarded as financial advice. We make every effort to keep the content accurate and current, but we cannot warrant its precision, completeness, or reliability. GNcrypto does not take responsibility for any mistakes, omissions, or financial losses resulting from reliance on this information. Any actions you take based on this content are done at your own risk. Always conduct independent research and seek guidance from a qualified specialist. For further details, please review our Terms, Privacy Policy and Disclaimers.
Articles by this author
