Tencent releases Hy4, open-source AI model for code, on Hugging Face

Tencent published a preview version of a new open-source AI model on Friday aimed at tasks such as software engineering, research, and financial analysis. The announcement was made in a post on Hugging Face, an open-source model repository, and reported by Reuters via ET Tech.

What Hy4 preview is

The model, called Hy4 preview, uses a mixture-of-experts (MoE) architecture with 770 billion parameters in total, according to Tencent, but only about 49 billion are used for each text request. This is the central feature of the announcement and the one that matters most to anyone thinking about inference cost.

In an MoE model, the total number of parameters defines the stored "capacity," but only a subset (the "experts" activated) processes each input. In practice, this means Hy4 carries the knowledge of a giant model, but the computational cost per request stays closer to that of a model with ~49B active parameters. It's the same design principle that other recent open models have been adopting to balance quality and cost.

Where Tencent plans to use it

The company said it plans to integrate the model into its own products, including CodeBuddy and WorkBuddy. In other words, besides making the weights available on Hugging Face, Tencent wants Hy4 running within its own ecosystem of productivity and coding-assistance tools.

Tencent itself acknowledged limitations of the early-stage model: according to the company, Hy4 preview can take longer than necessary to solve complex questions and sometimes over-verifies its own answers (over-verify). These are honest caveats for a release labeled as a preview, and they're worth considering before any production testing.

The context: Tencent accelerating in a crowded market

The release comes as Tencent increases its investment in AI in an increasingly competitive market. In April, the company introduced the Hunyuan 3.0 language model, its first major launch since hiring Yao Shunyu, a former OpenAI researcher, to lead the development of its AI platform.

Hy4 joins a wave of open models of Chinese origin that have started competing head-on with proprietary American models, especially from a cost angle. For the market, the message is clear: models capable of coding are no longer the exclusive domain of closed labs, nor billed only per API.

What changes for the Brazilian developer

For those building software in Brazil and testing generative models in production, the practical point is the same as always with open models: control and cost. A model with weights on Hugging Face can, in theory, be downloaded and run on in-house infrastructure or GPU providers, without depending on a pay-per-token API or sending proprietary code to third-party servers. This matters in scenarios involving sensitive data, compliance requirements, or tight budgets.

Hy4's MoE architecture reinforces this cost appeal: activating only ~49B parameters per request, instead of the full 770B, reduces the compute demand per call. Even so, hosting a model of this size is not trivial: the full 770B need to fit in memory, which requires robust hardware or multiple GPUs, and infrastructure costs can exceed those of simply using a ready-made API in many cases.

A few cautions before putting Hy4 into any serious pipeline:

• It's a preview. Tencent itself points to slowness on complex problems and excessive self-verification.
• License matters. "Open-source" in AI model jargon doesn't always mean a permissive license for commercial use; it's essential to read the terms published on Hugging Face before adopting it.
• Independent benchmarks. The available material doesn't include performance comparisons against competing code models; it's worth waiting for third-party evaluations.

What remains open

The announcement doesn't detail the commercial-use license, comparative public benchmarks, exact hardware requirements for inference, or a timeline for a stable version outside the preview label. There's also no confirmed availability of the CodeBuddy and WorkBuddy products outside China in the source. For a Brazilian team wanting to evaluate Hy4, the concrete next step is to check Tencent's repository on Hugging Face, review the license, and run its own tests on a representative codebase before any adoption decision.