Tencent releases open-source Hy4 preview with 770 billion parameters
The Apache-licensed mixture-of-experts model activates 49 billion parameters per token, supports a one-million-token context window and launches with API pricing below $1 per million input tokens.
Takeaways by Learning The World AI Show Hide
- Tencent has released Hy4 preview, a mixture-of-experts model with 770 billion total parameters, 49 billion active parameters and a one-million-token context window.
- The Apache 2.0 weights are available on Hugging Face, while hosted access costs $0.834 per million input tokens and $2.501 per million output tokens through OpenRouter.
- Tencent says Hy4 slightly beat GLM-5.3 and Kimi K3 in an internal blind test, but the result has not yet been independently reproduced.
AI-generated from this article and reviewed by the editor.
Tencent has released and open-sourced Hy4 preview, a large mixture-of-experts language model aimed at coding, office work and scientific research.
According to Tencent’s announcement on 28 August, the model has 770 billion parameters in its main network, activates 49 billion parameters for each token and supports a context window exceeding one million tokens.
The model’s Hugging Face card lists 78 layers, 256 routed experts and one shared expert in each mixture-of-experts layer. Each token activates eight routed experts alongside the shared expert. Tencent has released both the standard and FP8 weight sets under the Apache 2.0 licence.
That licence permits commercial use and modification, but the model remains demanding to run locally. The full Hugging Face repository is roughly 1.56 terabytes, and Tencent’s deployment instructions are designed around multi-GPU serving through vLLM or SGLang.
Tencent is presenting Hy4 preview as an open model for practical, long-running work rather than only benchmark questions. The company says it trained and evaluated the model with software engineers, game developers, finance analysts and security specialists working on coding, document, analysis and research tasks.
Its most prominent performance claim comes from an internal blind comparison involving 163 Tencent experts and 203 engineering tasks. Tencent says Hy4 preview averaged 2.99 points out of four, compared with 2.92 for GLM-5.3 and 2.94 for Kimi K3. Hy4 won 46.8% of comparisons against GLM-5.3 and 51.2% against Kimi K3.
Those results were produced by Tencent using its own task set and evaluators. They have not yet been reproduced by an independent benchmark organisation under a common public harness, so they show how the model performed in Tencent’s test rather than establishing a general ranking of the three systems.
Tencent also describes the release as an early preview with known limitations. The model card says Hy4 can spend longer than necessary reasoning through difficult tasks and may over-verify its own work. The company plans to use feedback from the preview before releasing further Hy4 models.
The weights are available through Hugging Face, ModelScope, GitCode and CNB. Hosted access is being added to Tencent’s WorkBuddy, CodeBuddy, Yuanbao and ima products, with two weeks of free use through WorkBuddy and CodeBuddy. Developers can also connect through Tencent Cloud TokenHub and OpenRouter.
OpenRouter lists Hy4 preview at $0.834 per million input tokens and $2.501 per million output tokens, with cached input priced at $0.042 per million tokens. The hosted version exposes the model’s one-million-token context window without requiring developers to operate its weight files themselves.
The release adds another very large Chinese-developed model to the growing open-weight market. Its practical appeal will depend on more than Tencent’s internal comparison: independent evaluations, serving requirements, output quality and the cost of successful tasks will determine how widely developers adopt it.
For now, the durable facts are the release itself: Apache-licensed weights are publicly available, the model exposes a million-token context window and commercial API access is live. Tencent’s claim that Hy4 has moved ahead of GLM-5.3 and Kimi K3 remains a first-party result awaiting independent testing.
