Tencent Open-Sources 770-Billion Parameter Hy4 Preview Model for Enterprise AI

On August 28, 2026, Tencent's Hunyuan team released the preview version of its new flagship language model, Hy4, placing a 770-billion parameter system into the open-source ecosystem under an Apache 2.0 license. The model, which activates only 49 billion parameters per token through a Mixture-of-Experts (MoE) architecture, represents a massive leap from the company's previous releases and is explicitly designed for heavy productivity workloads rather than casual conversation.[2][3][5]

The sheer scale of Hy4 preview pushes the boundaries of what open-weight models can currently handle. It features a one-million token context window, allowing it to ingest entire codebases, massive datasets, or hundreds of financial documents in a single prompt. This marks a nearly fourfold increase in context capacity compared to the 256,000-token limit of the previous Hy3 generation, enabling the model to manage long-horizon tasks without losing track of early instructions.[2][4]

Under the hood, the model utilizes 78 layers to process information. The first layer uses a standard dense feed-forward network, while the remaining 77 layers employ the MoE design. Each of these MoE layers contains 256 routed experts and one always-active shared expert, with the system routing every token to the top eight most relevant experts. This sparse activation keeps computational costs manageable during inference by only using roughly 6 percent of the network at any given time.[1][4]

Despite the efficient MoE routing, the 770-billion total parameter footprint means local deployment remains a significant hurdle. All 770 billion weights must reside in GPU memory simultaneously, meaning local hosting requires substantial enterprise hardware, such as a cluster of 16 B200 or 8 B300 GPUs. For most developers, this pushes the model out of the realm of local consumer hardware and into dedicated server racks or cloud-based API consumption.[1][2]

One of the most notable technical achievements of the Hy4 release is its role in its own development. According to technical documentation, Hy4 preview is the first Hunyuan model to actively participate in optimizing its own training pipeline and inference infrastructure. Rather than designing its own architecture from scratch, the model was used to independently locate and resolve bottlenecks in the reasoning system, achieving a measured 31.8 percent improvement in end-to-end throughput compared to the baseline.[1]

Tencent is positioning the model directly against the top tier of both open and closed-source competitors. In an internal blind evaluation involving 163 experts and 203 engineering tasks, Hy4 preview scored an average of 2.99 out of 4.00, edging out regional rivals like Kimi K3, which scored 2.94, and GLM-5.3, which scored 2.92. On public benchmarks, the model demonstrates strong competence in software engineering, hitting 85.4 points on Terminal-Bench 2.1 and 64.3 points on Deep-SWE.[3][4]

For developers unable to host the massive model locally, Tencent has made Hy4 preview available through its Cloud TokenHub and third-party routers like OpenRouter. The API pricing is set aggressively at $0.834 per million input tokens and $2.501 per million output tokens. While this undercuts several proprietary models in the same capability class, it remains significantly more expensive than the previous Hy3 generation, reflecting the massive increase in parameter count and context capacity.[2][5]

Despite the open-weight release, Tencent has maintained strict safety alignments on the model. The official API versions and the weights available on Hugging Face include rigorous filtering, meaning the model is not an uncensored system for unrestricted generation. The focus remains squarely on enterprise utility, with Tencent integrating the model into its own internal product suite—including WorkBuddy and Hunyuan Assistant—for a 127-day validation cycle covering office workflows and long-context stability before the public release.[2][3]

The provided technical documentation and reviews from Miraflow AI and MindStudio do not contain direct quotations from Tencent executives regarding the release. However, the launch of Hy4 preview establishes a new baseline for what open-weight models can achieve in enterprise environments. As developers begin integrating the one-million token context window into agentic coding workflows and scientific research pipelines, the primary bottleneck shifts from model availability to hardware accessibility. The next verifiable milestone for the Hunyuan team is the transition from this preview build to a final production release, which will test whether the 31.8 percent throughput gains hold up under global API traffic.[1][3]