Tencent's Hy3 removes licensing barrier, but GLM-5.2 keeps coding crown
Tencent released Hy3, a 295-billion-parameter open-weight model under Apache 2.0 license, removing regional restrictions that previously blocked EU, UK, and South Korean deployments. In blind testing against GLM-5.1, Hy3 scored 2.67 versus 2.51, with advantages in frontend development and CI/CD work. However, GLM-5.2 maintains a coding lead across benchmarks despite Hy3 operating at less than half the parameters.
TL;DR
- Tencent's Hy3 (295B parameters, 21B active) ships under Apache 2.0, removing prior geographic license restrictions that blocked EU, UK, and South Korean enterprise use
- Blind study of 270 experts across 312 comparisons showed Hy3 outperforming GLM-5.1 (2.67 vs 2.51), with strongest gains in frontend development, CI/CD, and data/storage tasks
- GLM-5.2 retains coding leadership on SWE-bench Verified (84.2 vs 78.0), SWE-bench Multilingual (83.0 vs 75.8), and Terminal-Bench 2.1 (81 vs 71.7)
- Model developed over 10 weeks post-preview, incorporating feedback from 50 internal product teams; includes 3.8B-parameter multi-token prediction layer and 256K context window
Why It Matters
License restrictions have been a hidden constraint on enterprise adoption of Chinese open-weight models. Tencent's Apache 2.0 release removes a legal barrier that previously forced companies serving EU, UK, or South Korean traffic to reject otherwise competitive models. This shifts the competitive calculus for enterprises evaluating open alternatives to proprietary APIs.
Business Impact
For IT teams and enterprises, Hy3 at half the parameters of GLM-5.2 offers a deployment economics argument: comparable performance on most tasks with lower compute costs and no geographic licensing friction. The model is free on OpenRouter for two weeks, lowering evaluation barriers. However, coding-focused teams should note GLM-5.2's persistent advantage on agentic coding benchmarks.
Key Implications
- Geographic license restrictions are now a material competitive disadvantage in open-weight model adoption, incentivizing permissive licensing for future releases
- Parameter efficiency messaging is becoming central to model positioning, with Tencent explicitly framing Hy3 as rivaling models with 2-5x the parameters
- Coding remains a distinct capability tier, with GLM-5.2 maintaining clear separation despite Hy3's gains elsewhere, suggesting specialized fine-tuning or architecture choices matter for agentic coding tasks
- Blind human evaluation on real-world workflows is emerging as a counter-narrative to public benchmarks, though results depend on which prior model version is tested
What to Watch
Monitor whether Hy3's coding performance improves in future updates, as the gap to GLM-5.2 is material for developer-focused deployments. Track adoption rates on OpenRouter and enterprise deployments post-free-tier period to assess whether license removal translates to actual usage. Watch for similar license changes from other Chinese model providers facing the same regional restrictions.
Subscribe to the newsletter
The latest stories and analysis, delivered to your inbox.
Free. No spam. Unsubscribe any time.
