Tencent’s Apache‑Licensed Hy3 Takes on GLM‑5.2 with Half the Size — Wins in Search, Loses in Code
### Tencent’s Open‑Source Leap: Hy3 Challenges Industry Titans with Lean Power Tencent has stripped away legal uncertainties that have long shadowed Chinese AI offerings by open‑sourcing its Hy3 model under the permissive Apache 2.0 license. The 295‑billion‑parameter Mixture‑of‑Experts (MoE) architecture activates just 21 billion parameters per token, delivering enterprise‑grade performance that rivals, and in some cases surpasses, larger competitors. Early benchmarks position Hy3 against Baidu’s GLM‑5.2, showing a clear advantage in search‑related tasks while lagging behind in code generation. #### Key Takeaways - **Open‑source breakthrough:** Hy3 is released under Apache 2.0, the most permissive license in the AI space, encouraging broad adoption and community contributions. - **Efficient MoE design:** Although the model contains 295 B parameters, only 21 B are active per token, cutting inference costs dramatically. - **Enterprise‑level metrics:** Hy3 matches or exceeds larger models on latency, throughput, and accuracy for several business‑critical workloads. - **Search superiority:** In head‑to‑head tests, Hy3 outperforms GLM‑5.2 on relevance and retrieval speed, positioning it as a strong candidate for enterprise search solutions. - **Code generation gap:** GLM‑5.2 retains an edge in programming‑related tasks, indicating room for further fine‑tuning of Hy3’s code‑centric capabilities. - **Strategic licensing impact:** The Apache 2.0 release removes barriers for integration into commercial products, potentially reshaping the competitive landscape of Chinese AI. - **Scalability promise:** The MoE approach enables future scaling without proportional increases in compute, aligning with cost‑effective AI deployment strategies. - **Regulatory clarity:** By adopting a globally recognized license, Tencent signals compliance and transparency, addressing longstanding concerns from international partners. - **Community engagement:** Open‑source availability invites contributions that could accelerate improvements in underperforming domains such as code synthesis. - **Market implications:** Hy3’s performance‑to‑size ratio may pressure rivals to adopt similar efficient architectures or reconsider licensing models. #TencentAI #Hy3 #ApacheLicense #MixtureOfExperts #EnterpriseAI #GLM5_2 #SearchAI #CodeGeneration #OpenModel #newsababil360 [Read Full Article](https://news.ababil360.com/tencents-apache-licensed-hy3-takes-on-glm-5-2-with-half-the-size-wins-in-search-loses-in-code/)
