AnswersPublished20 sources
Tencent Hunyuan Hy4 Preview Targets Coding, Research, and Long-Context Productivity
Tencent released Hy4 preview on August 28, 2026, as an open source mixture of experts model with 770 billion total parameters, 49 billion activated per request, and a context window exceeding one million tokens. Hy4 is designed less as a standalone chatbot than as the model layer behind coding, research, financial a...
Published byEdited with GPT-5.6 LunaImages generated with GPT Image 1.5
Research answer
Key takeaways
- Tencent released Hy4 preview on August 28, 2026, as an open source mixture of experts model with 770 billion total parameters, 49 billion activated per request, and a context window exceeding one million tokens.
- Hy4 is designed less as a standalone chatbot than as the model layer behind coding, research, financial analysis, document, and other productivity workflows in products such as WorkBuddy and CodeBuddy.
- Users can access it through Tencent products, Tencent Cloud TokenHub, and OpenRouter; reported international TokenHub pricing is $0.834 per million input tokens, $2.501 per million output tokens, and $0.042 per millio...
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What is Tencent Hunyuan’s Hy4 preview, released and open-sourced on August 28, 2026, and how do its 770 billion total parameters, 49 billion. Article summary: Tencent Hunyuan Hy4 preview is Tencent’s open-source, productivity-oriented next-generation language model, released on August 28, 2026. It is a large mixture-of-experts-style model with 770 billion total parameters, 49 . Topic tags: general, news, general web, user generated. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts w
Tencent Hunyuan Hy4 preview is Tencent’s open-source, productivity-focused large language model, released on August 28, 2026. Its headline specifications are substantial: a mixture-of-experts design with 770 billion total parameters, about 49 billion activated for a given text request, and a context window exceeding one million tokens. Tencent is positioning it for coding, research, financial analysis, and broader knowledge-work workflows rather than as a general-purpose model release alone. 1
3
14
The important qualification is that Hy4’s early performance advantage is modest and comes primarily from Tencent’s own evaluation. The model looks significant because of its scale, long context, product integration, and low reported usage cost—not because the available evidence proves that it broadly outperforms every competing open-source model.
What makes Hy4 preview different?
Hy4 preview uses a mixture-of-experts approach. The 770 billion figure describes the model’s total parameter count, while roughly 49 billion parameters are activated for an individual text request. That design allows Tencent to build a very large model while limiting the portion used for each request, although total deployment cost and real-world speed still depend on the surrounding inference system. 1
The model also supports a context window of more than one million tokens. In practical terms, that is aimed at workflows involving large codebases, multiple documents, long research materials, or cross-file analysis. A large context window does not automatically guarantee accurate retrieval or reasoning across every document, but it gives Hy4 a foundation for tasks that exceed the limits of shorter-context systems. 3
9
Tencent describes Hy4 as a top-tier open-source model with expanded model scale, context length, and training data. Those are company positioning claims; the available launch evidence supports the specifications and intended use cases more clearly than it establishes a universal ranking. 5
6
Coding and productivity are the main use cases
Tencent’s stated target areas include:
- Software engineering and agentic coding
- Front-end engineering
- Office and knowledge-work tasks
- Financial analysis
- Scientific research
- Game development
- Cross-file and data analysis
Reuters independently described the model as targeting software engineering, research, and financial-analysis work. Tencent’s own materials and launch coverage describe a broader productivity scope, including office work, coding, and science. 1
6
9
That positioning matters because Hy4 is being developed alongside the products intended to use it. Tencent says it is co-designing the model with WorkBuddy and CodeBuddy, using product workflows to optimize agentic coding and other productivity scenarios. This model–product relationship gives Tencent a way to test capabilities in practical software and knowledge-work environments instead of treating the model as an isolated API. 2
6
14
How Hy4 compares with GLM 5.3 and Kimi K3
Tencent evaluated Hy4 preview in a blind side-by-side test involving 163 internal experts and 203 engineering tasks. The reported average scores were:
- Hy4 preview: 2.99 out of 4.00
- Kimi K3: 2.94 out of 4.00
- GLM 5.3: 2.92 out of 4.00
Hy4 therefore finished slightly ahead of both models in this evaluation. The model’s reported win rates were 46.8% against GLM 5.3 and 51.2% against Kimi K3, with ties and losses making up the remainder. 5
30
The result should be read as an early signal, not a definitive industry ranking. The test used Tencent experts and engineering tasks, and the score differences were small. It does not establish that Hy4 will perform better across independent benchmarks, unrelated domains, or every real-world workload. The strongest defensible conclusion is that Tencent’s preview model performed competitively—and narrowly led its two named rivals in Tencent’s internal test. 5
30
Tencent says Hy4 helped optimize its own development process
One of the more unusual claims around Hy4 is that the preview model participated in optimizing parts of its own development pipeline. Tencent says it used Hy4 to help improve training methods, data strategy, evaluation processes, operators, and inference systems. That represents an early form of recursive self-improvement: the model is used to propose or assess changes to the systems used to develop and run models. 14
Tencent also reports that model-guided analysis of inference bottlenecks led to multiple rounds of work on operator fusion and communication efficiency. The company says this increased end-to-end inference throughput by 31.8% compared with its baseline. That is a reported engineering result from Tencent, not an independently audited replication, so it should be treated as evidence of the company’s development approach rather than a universally verified performance figure. 5
6
Where to use Hy4 preview
Hy4 preview is available as an open-source model and through Tencent’s product ecosystem. Launch coverage lists access through:
- WorkBuddy
- CodeBuddy
- Yuanbao
- ima
- Tencent Cloud TokenHub
- OpenRouter
WorkBuddy and CodeBuddy offered a two-week free access period at launch, while developers could connect to the model through TokenHub or OpenRouter. Availability, quotas, and product access may differ by region and platform. 5
9
14
Reported API pricing
The reported international TokenHub prices are:
- Input: $0.834 per million tokens
- Output: $2.501 per million tokens
- Cached input: $0.042 per million tokens
These figures make cached, repeated context particularly relevant for applications that repeatedly send the same codebase or document set. Actual project cost will still depend on prompt size, output length, caching behavior, concurrency, and the provider used to serve the model. 11
14
Why WorkBuddy is central to the launch
Hy4 preview is part of a broader Tencent strategy to pair model development with a productivity application. WorkBuddy, launched in March 2026 according to the supplied launch context, is intended to put Hunyuan capabilities into everyday engineering and knowledge-work tasks. Tencent and some coverage have framed WorkBuddy as China’s answer to Codex, but that is positioning language rather than an independently verified market ranking. 3
The strategy also helps explain the launch-day attention. Reports said demand created task-queue backlogs after Hy4 was integrated into WorkBuddy, prompting Tencent to expand inference capacity. That early congestion demonstrates interest, but it is not a measure of model quality by itself. 4
Tencent’s release cadence is another part of the story. Coverage reported that, after rebuilding its infrastructure in February, Hunyuan had been iterating major versions roughly every two months; Hy4 was described as the company’s third major model release in about six months. Tencent has also said that another Hy4 version is planned soon, although the available evidence does not establish a precise release date or final specifications. 2
3
The bottom line
Hy4 preview’s significance comes from the combination of four elements: unusually large total model scale, more than one million tokens of context, a direct focus on productivity agents, and open availability through Tencent’s products and API partners. Its internal score of 2.99 out of 4.00 narrowly exceeded GLM 5.3 and Kimi K3, but that result should not be presented as proof of broad superiority.
For developers, the most practical question is whether Hy4’s long-context handling, coding performance, pricing, and integration with WorkBuddy or CodeBuddy improve a specific workflow. For researchers and buyers, the next important evidence will be independent testing, clearer documentation of deployment behavior, and performance reports outside Tencent’s internal task set.
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "Tencent Hunyuan Hy4 Preview Targets Coding, Research, and Long-Context Productivity"?
Tencent released Hy4 preview on August 28, 2026, as an open source mixture of experts model with 770 billion total parameters, 49 billion activated per request, and a context window exceeding one million tokens.
What are the key points to validate first?
Tencent released Hy4 preview on August 28, 2026, as an open source mixture of experts model with 770 billion total parameters, 49 billion activated per request, and a context window exceeding one million tokens. Hy4 is designed less as a standalone chatbot than as the model layer behind coding, research, financial analysis, document, and other productivity workflows in products such as WorkBuddy and CodeBuddy.
What should I do next in practice?
Users can access it through Tencent products, Tencent Cloud TokenHub, and OpenRouter; reported international TokenHub pricing is $0.834 per million input tokens, $2.501 per million output tokens, and $0.042 per millio...
Sources
reuters.comChina's Tencent releases new open-source AI model for coding, research tasks
news.futunn.comQueue exceeds 5,000! Yao Shunyu's 'new project,' Hunyuan Hy4 Preview, goes viral upon launch
tencent.comTencent Releases and Open-Sources Tencent Hy4 preview
technode.comTencent open-sources Hy4 preview with 770B parameters and a 1M-token context · TechNode