Wednesday, September 2, 2026

Today’s Edition

AI Intel Report

MARKETS

Frontier Models

Tencent Hy4 Preview Marks Fastest Scale-Up in Open-Weight LLMs

The 770B-parameter model with 1M context window arrives 53 days after Hy3, boosting capabilities in coding and productivity while remaining fully open under Apache 2.0.

5 MIN READ
Inside a vast Tencent-operated data center facility in a modern industrial complex in southern China, multiple anonymous technicians wearing neutral-colored lab coats and business-casual attire work methodically among endless rows of tall black server racks equipped with dense GPU clusters and advanced liquid cooling pipes. One technician kneels at the base of a rack adjusting heavy power distribution units while another stands on a small step stool inserting large rectangular accelerator cards into open slots, their movements precise and coordinated. In the midground, a third technician reviews diagnostic readouts on a laptop connected via thick cables to the hardware, surrounded by blinking status LEDs in blue and green hues across the rack faces. The background reveals additional parallel aisles of identical server infrastructure stretching into the distance under bright overhead industrial lighting, with visible HVAC ventilation ducts and fire suppression systems mounted along the ceiling. Scattered on a nearby workbench are generic toolkits, spare cooling fans, and bundles of network cables, emphasizing the rapid hardware scale-up process. The overall environment conveys intense focus on expanding computational capacity for large-scale open-weight model training and inference, with productivity-oriented workspaces visible through glass partitions where additional anonymous staff sit at desks using standard office computers. The scene captures the tangible infrastructure behind swift model iterations arriving weeks apart, including support for extensive context handling and specialized applications in software development and task automation tools, all within a clean, climate-controlled, high-security facility featuring concrete floors, safety signage without lettering, and distant views of loading docks for equipment transport. Every detail from the texture of the metal rack panels to the arrangement of cooling hoses and the positioning of workers underscores the industrial reality of deploying massive parameter models under open licensing frameworks hosted on platforms for community access and integration with coding assistants and workflow optimization software.
Illustration: AI Intel Report

Tencent Hy4 preview is a Mixture-of-Experts large language model with 770 billion total parameters, 49 billion active per token, and a 1 million token context window, released as open weights under the Apache 2.0 license.

The launch of the Hy4 preview by Tencent on August 28, 2026, signals a notable acceleration in the development cycle of open-weight large language models from the company.

This release comes just 53 days after the Hy3 model and demonstrates substantial increases in both model size and context capacity.

The model aims to compete with closed frontier models in practical applications such as coding and productivity tasks by providing competitive performance in real-world scenarios.

What scale increases distinguish the Hy4 preview from the earlier Hy3 model?

Hy3, which was officially released on July 6, 2026, contained 295 billion total parameters with 21 billion active parameters and supported a context window of 256,000 tokens.

In comparison, the Hy4 preview expands to 770 billion total parameters with 49 billion active parameters and extends the context window to 1 million tokens.

This represents approximately a 2.6 times increase in overall size and nearly a 4 times expansion in context length over the course of 53 days.

The increase in active parameters from 21 billion to 49 billion supports more complex reasoning capabilities.

The context expansion allows the model to process entire code repositories or lengthy research papers in a single pass.

What technical features define the Hy4 preview model architecture?

The Hy4 preview is described as a new-generation Mixture-of-Experts flagship model.

It comprises 770 billion total parameters, of which 49 billion are activated per token.

The extended context window of 1 million tokens enables handling of very long inputs, which is particularly useful for tasks involving extensive documents or complex codebases.

The model has been co-designed with Tencent experts and products to target specific high-value use cases.

Weights for the model are available on platforms including Hugging Face, ModelScope, GitCode, and CNB.

The release is under the Apache 2.0 license, which permits broad commercial and research use with minimal restrictions.

Comparison of key specifications between Hy3 and Hy4 preview models.
ModelTotal ParametersActive ParametersContext WindowRelease Date
Hy3295 billion21 billion256K tokensJuly 6, 2026
Hy4 preview770 billion49 billion1M tokensAugust 28, 2026

How does the Hy4 preview perform in expert evaluations on engineering tasks?

In an internal blind evaluation involving 163 experts assessing 203 engineering tasks, the Hy4 preview achieved an average score of 2.99 out of 4.00.

This placed it ahead of GLM-5.3, which scored 2.92, and Kimi K3, which scored 2.94.

The evaluation used side-by-side comparisons where Hy4 preview recorded 46.8 percent wins against GLM-5.3 with 12.8 percent ties and 40.4 percent losses, and similar metrics against Kimi K3.

The blind evaluation setup with 163 experts provides a robust measure of real-world utility.

The win rates indicate consistent outperformance in the majority of cases.

Today we're releasing Hy4 preview, our most capable model to date. It's a 770B-parameter model with 49B active parameters and a 1M-token context window, and it makes the biggest gains where the work is hardest: long-horizon software engineering, document-heavy office work, and scientific research.Tencent Hy Team, Official announcement

What are the market and stakeholder implications of this open release?

The open-sourcing of such a large model under a permissive license could influence the competitive landscape by providing an accessible alternative to proprietary frontier models.

Developers and organizations may leverage the weights for customization in applications like game development and scientific research.

Platforms such as OpenRouter and tools like CodeBuddy and WorkBuddy could integrate the model to enhance their offerings in coding assistance and productivity.

Stakeholders in the AI ecosystem, including researchers and enterprises, stand to benefit from the transparency and reproducibility enabled by open weights.

The rapid iteration from Hy3 to Hy4 suggests that open models can keep pace with closed ones in terms of capability growth.

What are the primary optimization areas for the Hy4 preview in priority order?

  1. Long-horizon software engineering
  2. Document-heavy office work
  3. Game development
  4. Scientific research

What reactions and future directions are anticipated following the Hy4 preview release?

The official announcement emphasizes the model's strengths in challenging areas of work.

With availability on multiple repositories, adoption rates will be monitored closely by the community.

Future iterations may build upon this foundation to further close gaps with closed models in additional domains.

The co-design approach with internal products indicates a focus on practical utility.

This could lead to specialized fine-tunes or integrations that address specific industry needs over the coming months.

Continued development is expected to focus on further refinements based on community feedback.

The co-design with Tencent products suggests ongoing alignment with practical needs.

Future models in the series may continue the trend of rapid scaling.

Frequently asked

What is the parameter count of the Hy4 preview?

The Hy4 preview has 770 billion total parameters with 49 billion active per token.

When was the Hy4 preview released?

Tencent released the Hy4 preview on August 28, 2026.

How does Hy4 preview compare to Hy3?

Hy4 is approximately 2.6 times larger in parameters and has nearly 4 times the context window compared to Hy3.

Sources

  1. Tencent Hy — The model details, quote, and evaluation results for Hy4 preview.
  2. Tencent — The announcement of the release and open-sourcing of the Hy4 preview model with its specifications.
  3. Hugging Face — Hy4 preview is a new-generation Mixture-of-Experts (MoE) flagship model developed by the Tencent Hy Team. The model comprises 770B total parameters, of which 49B are activated per token.