Tencent’s latest release, Hy3, is making waves in the open-weight model community. With its 295-billion-parameter Mixture-of-Experts (MoE) model, Hy3 has been launched under the Apache 2.0 license, removing previous geographic restrictions that barred its use in the EU, UK, and South Korea. This move could position Tencent as a serious contender in the open-source AI space, especially with its promise of free access on OpenRouter for the first two weeks.
### From Preview to Product in Ten Weeks
Hy3’s journey from its initial preview in April to its full release has been remarkably swift, taking just ten weeks. This rapid development was driven by feedback from 50 internal product teams, which helped refine the model’s task execution and interaction capabilities. The architecture remains consistent with its preview version, featuring 295 billion total parameters and 21 billion active parameters per forward pass, thanks to its top-8 routing across 192 experts. The model also includes a 3.8-billion-parameter multi-token prediction layer for speculative decoding and a 256K context window.
Tencent’s focus with Hy3 is on reliability metrics and deployment economics, aiming squarely at production use. The company claims that the full release significantly outperforms similar-sized models and rivals flagship open-source models with two to five times the parameters. This claim sets the stage for a direct comparison with other leading models, particularly in the coding domain where GLM-5.2 currently holds the crown.
### Tencent’s Blind Test and Competitive Context
Tencent’s evaluation of Hy3 leans on a blind human study rather than standard public benchmarks. The study, involving 270 experts across various disciplines, yielded 312 valid comparisons, with Hy3 scoring an average of 2.67 out of 4 against GLM-5.1’s 2.51. The model demonstrated strong performance in frontend development, CI/CD, and data and storage tasks.
However, this evaluation targeted the older GLM-5.1 model, not the latest GLM-5.2, which was released by Zhipu AI in mid-June. According to Tencent’s benchmark appendix, GLM-5.2 outperforms Hy3 in coding tasks, including the SWE-bench Verified, SWE-bench Multilingual, Terminal-Bench 2.1, and DeepSWE. The results underscore GLM-5.2’s dominance in the coding domain, driven by its larger parameter size.
### Implications for Founders, Engineers, and the Industry
For founders and engineers, the release of Hy3 under the Apache 2.0 license means fewer legal hurdles and broader applicability, especially for enterprises operating in regions previously restricted by license terms. The model’s focus on reliability and deployment economics makes it an attractive option for production environments, providing a viable alternative to more resource-intensive models.
However, for those in the coding space, GLM-5.2 remains the benchmark, and Tencent’s Hy3 might not yet be the go-to model. Engineers aiming to leverage open-weight models for coding might still find GLM-5.2 more suitable, given its superior performance in coding benchmarks.
As the open-weight model landscape evolves, the choice between models like Hy3 and GLM-5.2 will hinge on specific use cases and deployment needs. Tencent’s move to make Hy3 more accessible could spur further competition and innovation, benefiting the broader AI community.
### What Happens Next
Tencent’s release of Hy3 under a permissive license is a strategic move that could reshape the open-source AI landscape. As other companies respond, the focus will likely shift to optimizing models for specific applications and reducing deployment costs. For founders and engineers, this means keeping an eye on licensing terms and performance metrics as they navigate the growing array of open-weight models available for enterprise use.
