Tencent’s Hunyuan team has just redefined the open-weight model landscape with the release of Hy3, a 295-billion-parameter Mixture-of-Experts (MoE) model featuring 21 billion active parameters. Unlike its predecessors, Hy3 ships under the permissive Apache 2.0 license, removing legal barriers that previously restricted deployments in key markets like the European Union, the United Kingdom, and South Korea. This shift has sparked immediate excitement in the open-model community, with many researchers calling the licensing change the most significant development in years.
Tencent is offering Hy3 for free on OpenRouter for two weeks, signaling its confidence in the model’s capabilities. While its technical specifications remain unchanged from the April preview—295B total parameters, 21B active per forward pass via top-8 routing across 192 experts, and a 256K context window—the behavior has been refined through extensive feedback. Chief AI Scientist Shunyu Yao emphasized that the team collected insights from over 50 internal product teams, addressing issues in task execution and scaling up post-training pipelines.
From preview to polished product in ten weeks
The evolution from Hy3’s April preview to its full release was driven by real-world testing. Tencent’s approach was unconventional: instead of relying solely on automated benchmarks, the team prioritized user feedback to refine the model’s reliability and interaction quality. This iterative process led to measurable improvements, including a significant reduction in hallucination rates and commonsense errors.
Internal evaluations revealed that Hy3’s hallucination rate dropped from 12.5% in the preview version to just 5.4% in the full release. Commonsense error rates fell from 25.4% to 12.7%, while multi-turn dialogue issues decreased from 17.4% to 7.9%. Tencent attributes these gains to stricter data cleaning protocols and training constraints designed to enforce grounded responses—answer only when evidence exists, state uncertainty clearly, and avoid fabricating information. The model’s performance on the open MRCR long-dialogue benchmark also improved dramatically, jumping from 42.9% to 75.1%.
How Hy3 compares to GLM-5.2 in coding and beyond
Tencent’s benchmarking strategy highlights Hy3’s strengths while acknowledging its limitations. The company conducted a blind human study involving 270 experts across disciplines, who completed 312 real-world workflow comparisons. Hy3 scored 2.67 out of 4 against GLM-5.1’s 2.51, with the most notable advantages in frontend development, CI/CD pipelines, and data storage workflows. However, the comparison is nuanced: GLM-5.2, released by Zhipu AI in mid-June, remains the coding leader, outperforming Hy3 in key benchmarks like 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).
The size disparity explains part of this gap. GLM-5.2 is a 744-billion-parameter MoE model with roughly 40 billion active parameters per token, while Hy3 operates with less than half the parameters and nearly half the per-token compute. Tencent’s positioning suggests Hy3 is optimized for workloads beyond pure coding—particularly in search, tool orchestration, and agentic tasks.
Hy3’s enterprise appeal: reliability, licensing, and deployment economics
For IT leaders evaluating open-weight models, Hy3’s Apache 2.0 license is a game-changer. Unlike many Chinese open models, which impose geographic restrictions, Hy3 is freely deployable worldwide, eliminating legal hurdles for global enterprises. Tencent’s focus on reliability metrics—such as reduced hallucination rates and consistent multi-turn performance—further strengthens its case for production use.
The model also leads in agentic search, achieving 84.2 on BrowseComp and 91.0 on DeepSearchQA, outperforming all other open models in Tencent’s table and rivaling proprietary systems like Claude Opus 4.8 and GPT-5.5. On tool orchestration, Hy3 scores 79.1 on the public MCP-Atlas set, while its performance on long-context retrieval (73.4 on AA-LCR) and agent-harness evaluations like ClawEval positions it as a top choice for search-and-tool-heavy workloads. Coding, however, remains GLM-5.2’s domain.
What’s next for Hy3 and the open-model ecosystem?
Tencent’s release underscores a broader trend: open-weight models are maturing beyond academic benchmarks and into practical, enterprise-ready tools. The Apache 2.0 licensing of Hy3 removes a critical friction point for global adoption, while its reliability improvements address long-standing concerns about hallucinations and consistency.
Independent verification of Hy3’s benchmarks is still pending, as most competitor numbers in Tencent’s appendix are based on the company’s own test runs. However, the model’s licensing flexibility and production-focused metrics suggest it could quickly gain traction among businesses prioritizing compliance, cost efficiency, and dependable performance. For now, the open-model race has a new frontrunner—one that prioritizes accessibility and reliability over sheer parameter count.
AI summary
Tencent’in Hy3 modeli, Apache 2.0 lisansıyla yayınlandı ve üretim odaklı yapısıyla dikkat çekiyor. Halüsinasyon oranları ve performans karşılaştırmaları hakkında detaylar.
