TL;DR: On August 28, 2026, Tencent released and open-sourced Hy4 preview, a 770-billion-parameter productivity model for coding, office work, and research.
Tencent Open-Sources Massive Hy4 Preview For Coding And Research
Tencent has released and open-sourced a new large language model built for hands-on work. The system, called Hy4 preview, was introduced on August 28, 2026, as a productivity-focused model covering software engineering, office tasks, game creation, and scientific research, according to Reuters and Tencent.
The launch centers on scale and practical use. Tencent describes Hy4 preview as a next-generation model with 770 billion total parameters, of which about 49 billion are active for any single request. It uses a mixture-of-experts design and supports a context window above 1 million tokens, a combination intended to handle long documents, large codebases, and multi-step analysis without losing track of earlier material.
Distribution is both open and product-integrated. The weights are available as an open-source release, including through a posting on Hugging Face, a widely used repository for open-source AI models. At the same time, global users can try the model inside WorkBuddy and CodeBuddy, as well as Yuanbao, ima and other Tencent products, or connect by API through Tencent Cloud TokenHub and OpenRouter.
The timing places Tencent more firmly in a crowded race to supply capable open models for developers and enterprises. Reuters reported that the company is increasing AI investment and noted that Hy4 preview followed the April introduction of Hunyuan 3.0, described as its first major release after bringing former OpenAI researcher Yao Shunyu in to lead AI platform development.
What Tencent Released On August 28
The August 28 introduction was framed as a preview rather than a final version. Tencent said it follows a preview-first method, gathering real-world feedback before issuing official releases, with additional models in the Hy4 series expected to roll out soon.
Tencent said Hy4 preview was enlarged substantially in model size, context length, and data volume compared with prior work, with progress in both pre-training and post-training producing a major step in overall capability. The company characterized the result as ranking among the top tier of open-source models.
The preview label also carries caveats. Tencent noted that the early-release system can sometimes take longer than needed on complex questions and may over-verify its own answers. That caution was highlighted in reporting by Reuters on the Hugging Face post.
For users, the immediate access terms include a trial window. Upon launch, Hy4 preview was offered free on WorkBuddy and CodeBuddy for two weeks. Free access to the earlier Hy3 on both platforms was extended until September 30, according to Tencent.
Scale, Design And Context Window
The headline technical figures are 770 billion total parameters and 49 billion active parameters. In a mixture-of-experts architecture, only a subset of the total system engages for each token or request, which is intended to preserve capability while limiting computation per query.
The context window exceeding 1 million tokens is aimed at long-context work. Tencent linked that capacity to software tasks that require understanding, planning, debugging, and validation across extended development sessions. It also pointed to office and analytical uses that involve multiple documents and lengthy workflows.
Tencent did not present Hy4 preview as a general chatbot alone. It described the model as built for real-world productivity, with training and product tuning oriented toward tasks people already perform in engineering teams, offices, studios, and laboratories.
Access routes reflect that positioning. Besides direct use in Tencent applications, developers can reach the model through cloud API channels. Published API pricing was set at USD 0.834 per million input tokens, USD 2.501 per million output tokens, and USD 0.042 per million tokens for cache hits, according to Tencent.
Internal Test Against Rival Chinese Models
To support its productivity claim, Tencent cited an internal blind evaluation focused on engineering work. The study involved 163 experts and 203 engineering tasks, according to Tencent.
In that exercise, Hy4 preview scored an average of 2.99 out of 4.00. Tencent compared that result with 2.92 for GLM-5.3 and 2.94 for Kimi K3. Reporting by The Latent identified those counterparts as Z.AI Co.'s GLM-5.3 and Moonshot AI's Kimi K3.
Tencent said the system was developed in close coordination with products including CodeBuddy and WorkBuddy, with attention to user experience in productivity settings. The company also said high-quality training data was co-created with its specialists in software engineering, gaming, finance, security, and other fields.
The comparison was conducted internally rather than by an outside benchmark body. Tencent presented it as evidence of strength on practical engineering tasks, not as an independent certification. The narrow numerical gap underscores that Chinese open models are competing closely on developer-oriented evaluations.
Coding, Office Work And Game Development Uses
In software engineering, Tencent said Hy4 preview offers stronger comprehension, planning, debugging, and validation for long-context projects. It also cited better visual quality and interaction experience for front-end development, an area where developers judge output directly in the browser.
For office productivity and analysis, the company pointed to improved understanding of complex work settings and stronger financial analysis. The model was optimized for data analysis and collaboration across documents, supporting a sequence from information processing to creation of documents, spreadsheets, and presentations.
Game development received particular emphasis. Tencent said Hy4 preview can produce a playable prototype from a single natural-language request and can work effectively with game engines. Developers can then continue refining elaborate game projects through multi-turn exchanges.
That workflow reflects Tencent's broader position across gaming, messaging, cloud, and workplace tools. Training input from specialists in gaming and software engineering, plus joint design with WorkBuddy, was intended to align the model with production routines rather than isolated demonstrations.
Research Capabilities And Self-Improvement Claims
For scientific work, Tencent said the model shows stronger ability to comprehend, reason through, and solve demanding research problems. It listed AI research and development, molecular dynamics simulation, condensed-matter physics, and fundamental mathematics among areas with notable improvement.
Two self-improvement claims stood out. Tencent said Hy4 preview participated for the first time in automated optimization of its own development process, covering training methods, data strategies, evaluation frameworks, and low-level operators. According to the company, the model suggested approaches, ran experiments, and iterated based on results, with code, logs, and feedback flowing into later rounds of exploration. Tencent called this an early-stage recursive self-improvement loop.
The second claim concerned inference efficiency. Tencent said the model independently examined bottlenecks in its inference system and performed multiple rounds of optimization involving operator fusion and communication optimization. Those changes raised end-to-end throughput by 31.8% compared with the baseline, with consistent gains across different context lengths and concurrency levels, according to Tencent. The company presented that as evidence of ability to optimize its own inference infrastructure autonomously.
Tencent also stressed cost efficiency as a route to wider use of advanced AI. The published token prices position Hy4 preview for high-volume coding assistance, document work, and research support through WorkBuddy, CodeBuddy, Yuanbao, ima, TokenHub, and OpenRouter, while the open-source release lets outside teams inspect and deploy the weights.
The next step, according to Tencent, is further iteration. Through its preview-first approach, the Hunyuan team plans to incorporate real-world feedback into research and development, with the next batch of Hy4-series models expected soon after the preview period.