Tencent launches Hy3 with Apache license and challenges GLM-5.2

Over the past year, the boom in open-weight models had an uncomfortable secret: many of the most powerful Chinese releases had licenses that excluded the European Union, the United Kingdom, and South Korea. This forced legal teams to block deployments before engineers even finished their evaluations, not only for companies headquartered in those regions but for any firm serving traffic to them. Tencent has just removed that obstacle with the launch of Hy3, a MoE model with 295 billion parameters (21B active) under the permissive Apache 2.0 license. The community reacted immediately: researchers on X pointed to the license change as the real headline, and a widely shared post argued that if the scores hold, Tencent has become one of the open-source leaders. Tencent is offering the model for free on OpenRouter for two weeks.

From preview to product in ten weeks

The Hy3 preview in April was the first model from Tencent's rebuilt pre-training and reinforcement learning infrastructure, launched less than three months after the rebuild in February. Chief AI scientist Shunyu Yao framed the early release as a deliberate move to gather feedback from developers and users before the official version. According to the model card on Hugging Face, the team collected feedback from over 50 product teams after the late April preview, fixed issues in task execution and interaction, and scaled their post-training pipeline. The architecture remains: 295B total, 21B active per forward pass via top-8 routing over 192 experts, a 3.8B MTP layer for speculative decoding, and a 256K context window. This rapid iteration approach, from preview to product in ten weeks, demonstrates a responsiveness to the community that few open-source models have achieved.

Performance: wins in human evaluation, loses in coding

Tencent's main evaluation is a blind study with 270 experts on real workflows. Hy3 scored 2.67 out of 4 versus 2.51 for GLM-5.1, with clear advantages in frontend, CI/CD, and storage. However, against GLM-5.2, the most recent version, Hy3 loses on all coding benchmarks: SWE-bench Verified (78.0 vs 84.2), SWE-bench Multilingual (75.8 vs 83.0), Terminal-Bench 2.1 (71.7 vs 81), and DeepSWE (28.0 vs 46.2). These figures, reported by VentureBeat, show that Hy3 is not the leader in coding but is competitive in general tasks. On other benchmarks like MMLU-Pro (80.0 vs 82.4 for GLM-5.2), GPQA Diamond (74.2 vs 75.8), and LiveCodeBench (74.0 vs 75.2), the differences are smaller. Notably, Hy3 achieves these results with only 21B active parameters compared to GLM-5.2's 29B, suggesting greater inference efficiency. Tencent's decision to prioritize reliability and deployment economy metrics over pure coding performance indicates a focus on enterprise production, where inference cost and latency are critical.

Implications for the open-source ecosystem

The Apache 2.0 license is the most significant change. Now companies worldwide can use Hy3 without fear of legal restrictions, a direct contrast to earlier Chinese models like Qwen2.5 or DeepSeek-V3, which used licenses that excluded certain regions. This positions Tencent as a serious player in open-source, directly competing with models like GLM-5.2 and Llama 3. The community has already reacted positively, highlighting the license as the real headline. Additionally, Tencent is offering Hy3 for free on OpenRouter for two weeks, facilitating mass testing. This move recalls Meta's release of Llama 2 in 2023, which also used a permissive license and catalyzed enterprise adoption. However, unlike Meta, Tencent is a Chinese company with its own product ecosystem (WeChat, Tencent Cloud), which could raise questions about long-term independence. Still, the license change removes geographic barriers, opening the model to global companies that previously avoided Chinese models due to legal risks.

What should readers know?

• Hy3 is ideal for general tasks and reasoning, but not the best choice for pure coding. For engineering teams prioritizing code generation, GLM-5.2 remains superior.
• The Apache 2.0 license removes geographic barriers, opening the model to global companies, including those in the EU, UK, and South Korea.
• Tencent has demonstrated rapid iteration capability: from preview to product in 10 weeks, with feedback from over 50 internal teams.
• GLM-5.2 remains the leader in coding, but Hy3 offers competitive performance with half the active parameters, reducing inference costs.
• The model is available for free on OpenRouter for two weeks, allowing developers and companies to test it without upfront investment.

"If the scores hold, Tencent becomes one of the open-source leaders," notes a researcher on X. This sentiment reflects Hy3's potential impact on the ecosystem, though questions remain about its long-term scalability and Tencent's independence.

In summary, Hy3 represents a milestone in the opening of Chinese models, but its success will depend on actual community adoption and whether Tencent maintains its commitment to open-source. Competition with GLM-5.2 and Llama 3 will be intense, but the Apache 2.0 license gives it a significant initial advantage.