Hy4 preview release and specifications
Tencent's Hunyuan unit released an open-source flagship large model called Hy4 preview on August 28, 2026, with a total of 770 billion parameters and 49 billion activated parameters, according to AIBase. The model supports a context length of 1 million tokens and is positioned for real productivity workloads, including software engineering, office tasks, and scientific research. Tencent described the release as an early iteration of the Hy4 family, with further pre-training and post-training improvements planned through agile iteration, and acknowledged known issues including long thinking on complex tasks.
Distribution channels and product integrations
Hy4 preview was simultaneously open-sourced on Hugging Face, GitHub, and ModelScope, and made available on Tencent Cloud TokenHub and OpenRouter, per AIBase. Users can access the model directly inside Tencent products including WorkBuddy/CodeBuddy, Yuanbao, and ima. Tencent's WorkBuddy unit separately launched an integration with Hy4 preview the same day, offering domestic and overseas versions with a two-week free trial and extending free access to the prior Hy3 model through September 30, as reported by finance.biggo.com.
Capability benchmarks and internal evaluation
Tencent said Hy4 preview was built with high-quality data co-developed alongside internal experts from software engineering, gaming, finance, and security, according to AIBase. In an internal blind test covering 203 engineering tasks scored by 163 experts, the model recorded an average score of 2.99 out of 4.00, compared with 2.92 for GLM-5.3 and 2.94 for Kimi K3. Tencent framed the strategy of releasing an early version openly as part of a broader industry shift from parameter-count competition toward productivity-effectiveness verification.
Targeted use cases
The release highlighted four productivity directions, per AIBase: improving understanding, planning, and debugging of long-horizon software-engineering tasks; completing end-to-end delivery from data analysis to documents, tables, and presentations in office scenarios; supporting game development through MCP integration with Unreal 5 and Unity, with the aim of generating playable demos from scratch via pure conversation; and covering scientific research areas including molecular dynamics, condensed matter physics, and basic mathematics.
Scientific research results tied to the model
Tencent reported two scientific results linked to Hy4 preview, according to AIBase. Combined with the Hyra model, it advanced the volume lower bound of the three-dimensional Blaschke–Lebesgue problem from 0.380799 to 0.41104, leaving roughly a 2% gap to the final proof of the Meissner tetrahedron conjecture. The combined setup also produced a 2.0x speedup in simulating a 32,512-atom phospholipid bilayer, reaching 54.9 ms per step.
Known limitations and next steps
Tencent stated that Hy4 preview is an early version of the Hy4 iteration and that pre-training and post-training still have room for improvement, with known issues including complex-task long thinking, AIBase reported. The company plans to absorb real feedback through agile iteration, and no further milestone dates were provided in the reporting.
Share this article
