Tencent's Apache-licensed Hy3 Takes on GLM-5.2 at Half the Size — Key

Tencent's Hunyuan team has released the full version of Hy3, a 295-billion-parameter Mixture-of-Experts (MoE) model, under the permissive Apache 2.0 license. This significant shift from its preview release in April removes previous regional restrictions that had hampered the adoption of Chinese models in key markets like the EU, UK, and South Korea. The move positions Hy3 as a major contender in the open-weight model landscape, challenging larger models on efficiency, reliability, and broad applicability for enterprise use.

The open-model community has reacted strongly, with researchers highlighting the license change as a pivotal moment. Many argue that if Hy3's performance holds up, Tencent could become a new leader in open-source AI. Tencent has also announced that Hy3 will be free on OpenRouter for two weeks, encouraging widespread adoption and testing.

From Preview to Production: Refining Hy3's Capabilities

Hy3's journey from an early April preview to its full release demonstrates a rapid development cycle, heavily influenced by user feedback. The model's initial preview marked the debut of Tencent's revamped pre-training and reinforcement learning infrastructure. Chief AI Scientist Shunyu Yao emphasized gathering developer feedback as a core reason for the early release.

Over ten weeks, the Hunyuan team integrated insights from more than 50 internal product teams. This feedback helped resolve issues in task execution and interaction, leading to a significantly scaled-up post-training pipeline. While the core architecture remains consistent—295 billion total parameters with 21 billion active per forward pass, facilitated by top-8 routing across 192 experts, a 3.8 billion-parameter multi-token prediction layer, and a 256K context window—the full release boasts improved behavior and performance.

Benchmarking Hy3: Where it Shines and Where it Trails

To showcase Hy3's capabilities, Tencent conducted a blind human study involving 270 experts working on real-world workflows. In these evaluations, Hy3 scored 2.67 out of 4 against the older GLM-5.1's 2.51, demonstrating clear advantages in frontend development, CI/CD, and data and storage tasks. This approach reflects Tencent's belief that public benchmarks don't always capture the full story of real-world utility.

However, a direct comparison with the newer GLM-5.2 reveals a different picture for coding-specific tasks. Tencent's own benchmark appendix shows GLM-5.2, a roughly 744-billion-parameter MoE with about 40 billion active parameters, maintaining its lead across agentic coding suites like SWE-bench Verified (84.2 vs. 78.0) and DeepSWE (46.2 vs. 28.0). Hy3, at less than half the parameters and per-token compute, concedes the coding crown to its larger rival.

Hy3 truly excels in other critical areas. It leads the open field in agentic search, scoring 84.2 on BrowseComp and 91.0 on DeepSearchQA, putting it on par with flagship proprietary models like Claude Opus 4.8 and GPT-5.5. Furthermore, Hy3 outperforms other open models in tool orchestration (79.1 on MCP-Atlas), agent-harness evaluations (ClawEval), and long-context retrieval (73.4 on AA-LCR). This suggests Hy3 is an optimal choice for search-and-tool-heavy agent workloads.

It's important to note that many competitor figures in Tencent's appendix are derived from their own internal test runs. Independent verification from indices like Artificial Analysis is still awaited, which will offer a third-party perspective on these claims.

Enterprise Focus: Reliability and Deployment Economics

Tencent's emphasis on reliability metrics and deployment economics signals a clear focus on enterprise adoption. The model card reads more like a production report, highlighting significant improvements in Hy3's stability and accuracy. Internal evaluations show the hallucination rate dropped from 12.5% in the preview to 5.4%, and commonsense error rates fell from 25.4% to 12.7%. These gains are attributed to meticulous data cleaning and training constraints designed to ensure grounded answers and prevent data fabrication.

Multi-turn behavior also saw substantial improvements, with issue rates on internal tests decreasing from 17.4% to 7.9%, and its score on the open MRCR long-dialogue benchmark jumping from 42.9% to 75.1%. Tencent also underscored Hy3's consistency across various agent scaffolds, an often-overlooked property crucial for enterprises that may use diverse agent frameworks. While these are self-reported internal measurements, the decision to highlight them indicates Tencent's understanding of what enterprise buyers value: models that perform reliably in production, not just in demos.

Economically, Hy3 presents a compelling case. Its 295 billion total parameters translate to an FP8 footprint of under 300GB, less than half the memory required by GLM-5.2 (approximately 744GB). This significantly reduces the hardware burden for self-hosting, potentially making the difference between needing multiple specialized nodes or a single, more attainable server. The model's design also aligns with geopolitical realities: Tencent recommends serving Hy3 on Nvidia's H20-3e, a GPU designed to comply with U.S. export restrictions to China. This constraint-driven sizing coincidentally makes Hy3 highly efficient on Western-available hardware like H100s, H200s, and B200s, facilitating standard vLLM and SGLang deployments.

With the Apache 2.0 license removing a major legal hurdle, Hy3 is poised to become a strong open-weight choice for reliability-sensitive applications and organizations seeking frontier-adjacent AI capabilities without the need for frontier-scale infrastructure. The market will now watch to see if Western enterprises embrace this Tencent offering or await further independent verification from sources like Artificial Analysis before committing.

FAQ

Q: What is the most significant change with Tencent's Hy3 release?

A: The most significant change is the adoption of the permissive Apache 2.0 license, which removes previous regional restrictions and makes Hy3 accessible to a much broader global enterprise audience, particularly in the EU, UK, and South Korea.

Q: Where does Hy3 excel compared to other open models, and where does it fall short?

A: Hy3 excels in agentic search, tool orchestration, agent-harness evaluations, and long-context retrieval, rivaling top proprietary models. However, it trails GLM-5.2 in agentic coding benchmarks, primarily due to GLM-5.2's significantly larger parameter count.

Q: Why is Hy3's smaller size and specific hardware recommendation important for enterprises?

A: Hy3's smaller size (295B parameters) translates to lower memory and compute requirements for deployment, making it more cost-effective and attainable for self-hosting. Its recommendation for Nvidia H20-3e GPUs, compliant with US export restrictions, ensures accessibility for Chinese companies, and conveniently, its optimized performance on these capped chips means it runs even more efficiently on readily available Western hardware.