Superpower Daily
Tencent Releases Open-Source Hy4, a 770B Model That Activates 4...Meta Tests Data-Center Repair Robots, but Cable Bots Still Need...Amap Releases ABot-Recon, a 12-Frame 3D Mapper for 10,000-Frame...Databricks Files $5B Equity Offering With Nearly All of It Alre...Snowflake Open-Sources Semi-Persistence With Sub-Second Single-...Anthropic-Backed Group Sends $500,000 to a Pro-Markey PAC as AI...Tencent Releases Open-Source Hy4, a 770B Model That Activates 4...Meta Tests Data-Center Repair Robots, but Cable Bots Still Need...Amap Releases ABot-Recon, a 12-Frame 3D Mapper for 10,000-Frame...Databricks Files $5B Equity Offering With Nearly All of It Alre...Snowflake Open-Sources Semi-Persistence With Sub-Second Single-...Anthropic-Backed Group Sends $500,000 to a Pro-Markey PAC as AI...
Modelspublished

Tencent Releases Open-Source Hy4, a 770B Model That Activates 49B Parameters Per Request

Tencent is pairing a large mixture-of-experts model with plans for product integration, while warning that the preview can spend too long working through difficult questions.

By 2 min read
Tencent Releases Open-Source Hy4, a 770B Model That Activates 49B Parameters Per Request
Sign in to save and react

Listen to this story

The audio brief

About 1:24
0:001:24
Read transcript
Tencent has released Hy4, an open-source preview that combines a headline-grabbing 770 billion parameters with a much smaller working footprint for each request. Hy4 uses a mixture-of-experts design: the full model contains about 770 billion parameters, but Tencent says roughly 49 billion are active when it answers any one text prompt. That distinction is central to the release. The model is aimed at software engineering, research, and financial analysis, rather than just one narrow coding task, and Tencent published the preview through Hugging Face on August 28. There is an important caveat. Tencent says Hy4 can take longer than necessary on difficult questions and may over-verify its own answers. For research and finance workflows, that behavior could become a meaningful part of the user experience, especially when speed matters as much as depth. The immediate product plan is to place Hy4 alongside Tencent’s existing CodeBuddy and WorkBuddy tools, although the company has not provided a deployment timetable. The release also extends Tencent’s recent model push, following its April unveiling of Hunyuan 3.0, after former OpenAI researcher Yao Shunyu took on a leadership role in AI platform development. The key thing to watch is whether Hy4’s large architecture and 49-billion-parameter activation translate into a useful product before its slow, over-verifying preview behavior becomes the limiting factor.

Story brief

3 key points

Tencent’s August 28 release gives developers an open-source preview of Hy4, a mixture-of-experts model whose 770B total parameters are intended to support demanding coding, research, and financial-analysis workloads without activating the full model on every text request. The claimed 49B active parameters may make deployment more tractable than the headline size suggests, but Tencent flags slower-than-necessary...

  1. 01

    Hy4 uses a mixture-of-experts design: 770B parameters overall, with approximately 49B active for each text request.

  2. 02

    Tencent warns that the preview may take too long on complex questions and over-verify answers.

  3. 03

    The model was released through Hugging Face as an open-source preview on August 28.

Tencent has released Hy4 preview, an open-source model aimed at software engineering, research, and financial analysis. Its design advertises a very large total model while using a smaller share of its parameters for each text request, and Tencent plans to put it alongside CodeBuddy and WorkBuddy.

One model, two parameter counts

Hy4 uses a mixture-of-experts architecture with 770 billion parameters in total. For any one text request, however, Tencent says approximately 49 billion parameters are active. The release therefore distinguishes the model’s overall parameter count from the portion used when it answers a prompt.

Tencent presented the preview in a post on Hugging Face, the repository for open-source AI models. The stated target tasks put the release squarely in technical and analytical workflows rather than a single-purpose coding product.

The architecture at a glance
770 billionTotal parameters

Tencent describes Hy4 as a mixture-of-experts model with 770 billion total parameters.

About 49 billionActive per text request

Tencent says about 49 billion parameters are used for a given text request.

A preview with a stated trade-off

Tencent also identifies a limitation in this early release: Hy4 can take longer than necessary when working through complex questions and may over-verify its own answers. That caveat is particularly relevant to the research and financial-analysis uses Tencent names, where a lengthy response process can become part of the product experience.

From model release to Tencent products

The immediate commercial path Tencent describes is integration alongside CodeBuddy and WorkBuddy. That is a plan, not a disclosed deployment timetable, but it places Hy4’s first identified product role inside Tencent’s existing software and workplace tools.

Tencent released Hy4 on August 28. The launch follows its April unveiling of Hunyuan 3.0, which was Tencent’s first major model release after former OpenAI researcher Yao Shunyu joined to lead AI platform development. Hy4 adds a new open-source preview to that model-release sequence, with its eventual integration and behavior beyond the disclosed early-release caveats still to be determined.

Sources

  1. wtvbam.comChina’s Tencent releases new open-source AI model for coding, research tasks
  2. asiaone.comTencent releases new open-source AI model for coding, research tasks

Related coverage

Continue reading