Back to list
Tencent Launches Hy4 Preview: A 770B Parameter Open-Source Model with 1M Token Context for Global Productivity
Product LaunchTencentOpen Source AILarge Language Models

Tencent Launches Hy4 Preview: A 770B Parameter Open-Source Model with 1M Token Context for Global Productivity

Tencent has officially released and open-sourced the Hy4 Preview, a next-generation large language model (LLM) designed to handle complex, real-world productivity tasks. Boasting a massive architecture of 770 billion total parameters and 49 billion active parameters, the model features a context window exceeding 1 million tokens. Developed through deep co-design with industry experts in fields such as software engineering, finance, and gaming, Hy4 Preview has demonstrated superior performance in coding, office work, and scientific research. In internal blind evaluations, it outperformed notable competitors like GLM-5.3 and Kimi K3. The model is now available globally via open-source channels, Tencent's productivity suite including WorkBuddy and CodeBuddy, and API platforms like Tencent Cloud TokenHub and OpenRouter, marking a significant advancement in the open-source AI landscape.

Hacker News
Share

Key Takeaways

  • Massive Scale and Efficiency: Hy4 Preview features 770 billion total parameters with 49 billion active parameters, utilizing an architecture designed for high-performance productivity.
  • Extensive Context Window: The model supports a context length exceeding 1 million tokens, allowing for the processing of vast amounts of data in a single session.
  • Expert-Validated Performance: In blind evaluations involving 163 experts, Hy4 Preview scored 2.99/4.00, surpassing competitors such as GLM-5.3 and Kimi K3.
  • Broad Accessibility: The model is open-sourced and integrated into Tencent products like WorkBuddy, CodeBuddy, Yuanbao, and ima, with API access via Tencent Cloud and OpenRouter.
  • Domain-Specific Optimization: Training data was co-created with experts in software engineering, gaming, finance, and security to ensure real-world utility.

In-Depth Analysis

Technical Architecture and Unprecedented Scale

Tencent's release of the Hy4 Preview represents a major leap in the evolution of open-source large language models. By implementing a design that utilizes 770 billion total parameters while maintaining 49 billion active parameters, Tencent has balanced raw power with operational efficiency. This approach allows the model to rank among the top tier of open-source offerings globally.

One of the most striking features of Hy4 Preview is its context window, which now exceeds 1 million tokens. This expansion is critical for modern productivity tasks, such as analyzing entire codebases, long-form legal documents, or comprehensive scientific research papers. The significant increase in model size, context length, and data volume compared to its predecessor, Hy3, indicates a strategic focus on overall intelligence and the ability to handle high-density information environments.

Productivity-Driven Development and Evaluation

Unlike models developed in isolation, Hy4 Preview was built through a process of deep co-design. Tencent collaborated with experts across diverse domains—including software engineering, gaming, finance, and security—to co-create high-quality training data. This ensures that the model is not just a general-purpose assistant but a specialized tool capable of handling professional-grade tasks.

To validate its effectiveness, Tencent conducted an internal blind evaluation. This rigorous test involved 163 experts and 203 specific engineering tasks. The results placed Hy4 Preview at an average score of 2.99 out of 4.00. This performance metric is particularly noteworthy as it edges out other prominent models in the industry; specifically, it outperformed GLM-5.3, which scored 2.92, and Kimi K3, which scored 2.94. These results suggest that the model's optimizations for coding and office work are yielding tangible advantages in real-world scenarios.

Ecosystem Integration and Global Availability

Tencent is ensuring that the power of Hy4 Preview is immediately accessible to both individual users and enterprise developers. The model has been integrated into Tencent’s core productivity applications, such as WorkBuddy and CodeBuddy, as well as consumer-facing products like Yuanbao and ima. For developers looking to build their own applications, the model is available via API through Tencent Cloud TokenHub and OpenRouter.

To encourage adoption, Tencent has launched a promotional period where Hy4 Preview is available for free on WorkBuddy and CodeBuddy for two weeks following its release. Additionally, the company has extended free access to the previous version, Hy3, until September 30. This strategy not only showcases the new model's capabilities but also maintains continuity for users within the Tencent AI ecosystem.

Industry Impact

The release of Hy4 Preview signals a shift in the open-source AI market toward models that are specifically tuned for professional productivity rather than general conversation. By open-sourcing a model of this scale (770B parameters) and capability, Tencent is challenging the dominance of proprietary models in the productivity sector.

The inclusion of a 1-million-token context window sets a new benchmark for open-source models, potentially accelerating innovation in fields that require long-context processing, such as automated software engineering and large-scale data analysis. Furthermore, the competitive performance against established models like GLM and Kimi highlights the rapid pace of development within the Chinese AI sector and its growing influence on the global open-source community.

Frequently Asked Questions

Question: What are the primary technical specifications of the Tencent Hy4 Preview?

Hy4 Preview features 770 billion total parameters and 49 billion active parameters. It also supports a context window that exceeds 1 million tokens, making it suitable for processing very large datasets and long documents.

Question: How does Hy4 Preview compare to other models like GLM-5.3 and Kimi K3?

In an internal blind evaluation conducted by Tencent involving 163 experts and 203 engineering tasks, Hy4 Preview achieved an average score of 2.99/4.00. This score is slightly higher than GLM-5.3 (2.92/4.00) and Kimi K3 (2.94/4.00), indicating superior performance in productivity-related tasks.

Question: Where can users and developers access the Hy4 Preview model?

Users can access the model through Tencent products such as WorkBuddy, CodeBuddy, Yuanbao, and ima. Developers can connect to the model via API through Tencent Cloud TokenHub and OpenRouter. It is also available as an open-source model for the broader community.

Read original article

Related News

OpenAI Introduces GPT-6 Sol and Luna Featuring Half API Pricing and Reduced Error Rates
Product Launch

OpenAI Introduces GPT-6 Sol and Luna Featuring Half API Pricing and Reduced Error Rates

OpenAI has officially introduced its newest model offerings, GPT-6 Sol and Luna, marking a notable shift in both performance and developer accessibility. According to reports, the new releases arrive at half the API cost compared to preceding options, significantly lowering the financial threshold for deploying advanced AI capabilities. Furthermore, internal testing indicates that GPT-6 Sol demonstrates substantial accuracy improvements, committing approximately half as many mistakes as its direct predecessor. This dual advancement—pairing dramatic cost reductions with superior reliability—positions the GPT-6 tier as a major development for builders, enterprise teams, and the broader artificial intelligence ecosystem seeking scalable and dependable model access without prohibitive compute expenditures.

Anthropic Unveils Claude Opus 5.5 with Lower Pricing Structure for Developers and Enterprise Workloads
Product Launch

Anthropic Unveils Claude Opus 5.5 with Lower Pricing Structure for Developers and Enterprise Workloads

Anthropic has officially unveiled Claude Opus 5.5, introducing a revised and lower pricing model for the model. According to reporting from Tech in Asia, the newly introduced tier sets access costs at US$4 per million input tokens and US$20 per million output tokens. This update highlights a defined 1:5 ratio between input consumption and output generation costs. By establishing explicit token-based rates, Anthropic positions Claude Opus 5.5 for broader commercial deployment across developer environments and enterprise API pipelines. While additional benchmark metrics and architectural specifications were not disclosed in the report, the announcement underscores a clear focus on lowering economic barriers for high-tier model utilization.

Product Launch

OpenAI Introduces Better Prompt Caching for GPT-6 Featuring Enhanced Diagnostics and Explicit Breakpoints

OpenAI has announced significant improvements to prompt caching for GPT-6 via an official OpenAI Blog update. The latest enhancements are designed to deliver higher cache hit rates while introducing new diagnostics, explicit breakpoints, and dedicated controls for developers. According to the announcement, these core prompt caching upgrades directly reduce latency and lower overall operational costs when running GPT-6 workloads. By providing explicit breakpoints and granular cache controls, the update gives developers enhanced mechanisms to optimize repeated prompt segments and track caching behavior effectively. This release reflects OpenAI's continued focus on performance optimization, cost reduction, and developer observability for GPT-6 deployments.