China’s Tencent has released Hy4 preview, an open-source large language model aimed at coding, office work, scientific research and other productivity tasks.
The company said in an August 28 announcement that the Mixture-of-Experts model has 770 billion total parameters, with 49 billion activated for each token, and supports a context window of more than 1 million tokens. The model’s Hugging Face card confirms a one-million-token context length and says the weights are available under the Apache License 2.0.
A larger open model for productivity
Hy4 preview is the successor to Tencent’s Hy3 model and marks a substantial increase in scale. Hy3 has 295 billion total parameters, 21 billion active parameters and a context length of up to 256,000 tokens, according to earlier TNGlobal coverage.
Tencent said Hy4 preview was developed with training data informed by its software engineers, game developers, finance analysts and security specialists. The company positioned the model for long-running software-development tasks, data analysis, document creation, cross-document collaboration, game prototyping and scientific problem-solving.
The model uses 78 backbone layers. Most employ a Mixture-of-Experts structure with 256 routed experts and one shared expert, while eight routed experts and the shared expert are activated for each token. Tencent also included a separate multi-token prediction layer for speculative decoding, according to the model card.
Hy4 preview can be accessed through Tencent products including WorkBuddy, CodeBuddy, Yuanbao and ima. Tencent also lists application programming interface access through Tencent Cloud TokenHub and OpenRouter. WorkBuddy and CodeBuddy offered the model free for two weeks from launch, while Tencent extended free access to Hy3 on those services through September 30.
Developers can download the full model and a lower-precision FP8 version. Tencent has also published instructions for serving the model with vLLM and SGLang, fine-tuning guidance and an AngelSlim compression toolkit. Those resources make the release usable beyond Tencent’s hosted products, although operators must still provide their own infrastructure and assess deployment costs.
Benchmarks remain company-reported
Tencent said 163 internal experts compared model outputs on 203 engineering tasks. Hy4 preview received an average score of 2.99 out of four, versus 2.92 for GLM-5.3 and 2.94 for Kimi K3.
Those results were produced through an internal evaluation and have not been independently verified. They also cover selected engineering tasks rather than the full range of workloads in which enterprises may use the model. The model card lists known limitations and recommends that users evaluate the system for their own applications.
Tencent also said Hy4 preview participated in parts of its own development process, including automated work on training methods, data strategies, evaluation frameworks and low-level operators. The company described this as an early-stage recursive self-improvement loop, but did not provide an independent assessment of the approach.
The open release gives developers access to the model weights, fine-tuning guidance, deployment instructions and quantization tooling. Its Apache 2.0 license permits broad commercial use, subject to the license terms, although the model’s size means self-hosting will require substantial computing and memory resources.
The launch adds another large Chinese model to a growing field of open-weight systems competing for developers and enterprise workloads. For Tencent, Hy4 preview also extends a product strategy that links its underlying model to workplace applications, cloud access and third-party developer platforms.
China’s Tencent makes AI model Hy3 available globally through WorkBuddy, cloud, developer platforms