Tencent Hy4 Preview: 770B Open-Source MoE Model Explained

Key Takeaways

• Tencent has released Hy4 preview, an open-source Mixture-of-Experts (MoE) flagship model.
• The model features 770 billion total parameters, with 49 billion active parameters.
• It supports a massive 1 million-token context window and is released under Apache 2.0 weights.
• The model is designed for high-productivity tasks, including coding and research.

A New Flagship in Open-Source AI

On August 28, 2026, Tencent introduced Hy4 preview, marking a pivotal moment in the democratization of frontier-scale artificial intelligence. Developed by the Tencent Hy Team, this next-generation large language model (LLM) is not merely an incremental update but a strategic leap in capacity and utility. The model is built upon a sophisticated Mixture-of-Experts (MoE) architecture, a design choice specifically engineered to handle the rigors of complex productivity tasks that typically exhaust smaller models.

Reuters reports that the model is specifically aimed at high-complexity tasks such as deep academic research and advanced software coding. By open-sourcing a model of this magnitude, Tencent is positioning itself as a primary contributor to the open-weight ecosystem, challenging the dominance of closed-source proprietary systems. The Hy4 preview is designed to serve as a versatile engine for developers and researchers who require the reasoning capabilities of a trillion-parameter class model without the prohibitive costs of traditional dense architectures.

Technical Specifications and Architecture

The Hy4 preview stands out due to its massive scale and highly efficient parameter management system. According to Tencent Hy, the model comprises a total of 770 billion parameters. In a traditional dense model, every single parameter would be activated for every single token processed, leading to astronomical computational requirements. However, Hy4 utilizes a Mixture-of-Experts (MoE) approach, meaning it utilizes only 49 billion active parameters during operation.

This MoE architecture allows the model to maintain a vast internal "knowledge base" across its 770B parameters while only engaging the most relevant "experts" (sub-networks) for any given query. This optimization ensures that the model maintains high cognitive capacity and nuance while significantly reducing the latency and hardware requirements for inference. Miraflow.ai notes that this architecture even helped the model optimize its own training and inference processes, creating a feedback loop of efficiency.

The Power of the 1-Million Token Context Window

One of the most significant technical milestones of the Hy4 preview is its expansive context window. Tencent Global notes that the model supports a context window of 1 million tokens or more. In practical terms, this allows the model to ingest, process, and remember vast amounts of information within a single session—equivalent to several long novels or an entire codebase of a medium-sized software project.

For researchers, this means the ability to upload dozens of scientific papers and ask the model to synthesize a literature review without losing the context of the first document. For programmers, it enables the model to understand the interdependencies of thousands of lines of code across multiple files. This "long-horizon work" capability, as highlighted by Product Hunt, transforms the model from a simple chatbot into a comprehensive analytical partner.

Performance Benchmarks and Competitive Edge

The release of Hy4 preview has already begun to shift the benchmarks for open-weight models. LinkedIn reports that the model scored 85.4 on TerminalBench, a critical metric for evaluating AI's ability to interact with terminal environments and execute complex coding tasks. Notably, this score indicates that Hy4 preview is beating DeepSeek V4 in specific technical domains, signaling a new peak in open-source coding proficiency.

Beyond coding, Gadinsider highlights that the model is equally potent in financial analysis and research. The combination of its MoE efficiency and its massive parameter count allows it to handle the nuanced logic required for financial forecasting and data synthesis, areas where smaller models often hallucinate or lose the thread of complex reasoning.

Availability, Accessibility, and Licensing

To facilitate widespread adoption and rapid community development, Tencent has adopted a highly permissive licensing strategy. Cellcog.ai notes that the model's weights have been released under the Apache 2.0 license. This is a critical detail for enterprises and independent developers, as it allows for the commercial use, modification, and distribution of the model without the restrictive clauses often found in "open-ish" AI licenses.

Tencent has ensured that the model is accessible through the industry's most vital hubs. Tencent Cloud confirms that the release includes direct links to GitHub and Hugging Face, where the model weights and implementation guides are hosted. For those who lack the massive hardware required to run a 770B parameter model locally, Mindstudio.ai suggests that the model can be run using optimized frameworks like vLLM or SGLang to manage memory more effectively.

Furthermore, for users who want to experience the model's capabilities without an initial infrastructure investment, Cellcog.ai mentions that it is available for a two-week free trial on CodeBuddy, providing a low-friction entry point for developers to test its coding prowess.

Market Positioning and Strategic Impact

The launch of Hy4 preview signals a strategic pivot by Tencent in the global AI arms race. While many companies are focusing on making models smaller and faster (the "small language model" or SLM trend), Tencent is proving that massive scale is still essential for frontier-level reasoning. Medium suggests that this release represents a shift in how major tech players approach AI deployment, moving toward "open-weight flagships" that can act as foundations for an entire ecosystem of fine-tuned derivatives.

Hy4 preview enters a crowded but high-stakes market, competing directly with other powerhouse models such as GLM 5.3, Qwen 3.8, and Kimi K3. By offering a model that combines the scale of 770B parameters with the efficiency of 49B active parameters and a massive context window, Tencent is attempting to set a new standard for what an open-source model can achieve. This move not only enhances Tencent's prestige in the AI community but also provides a powerful tool for the global developer community to push the boundaries of what is possible in automated research and software engineering.