Tencent Open Sources Hy3 Preview to Scale Efficient Agentic Reasoning

Tencent HyTencent Hy

· Updated

Tencent open-sourced Hy3 preview, a 295B-parameter model that activates only 21B parameters per token. The release signals a shift toward high-efficiency reasoning models designed specifically for autonomous agent workflows and complex reasoning tasks.

Tencent, a developer of foundation models for multimodal generation, open-sourced the weights for Hy3 preview, a large model optimized for reasoning and agentic tasks. This release extends controllable reasoning architectures that activate only 21B of 295B total parameters to maintain high inference efficiency.
Total parameters
295B
Active parameters
21B
Context window
256K tokens
Architecture
Mixture of Experts
Availability
Hugging Face, OpenRouter
Free trial period
Through May 8th 2026

The release mirrors the pattern seen in Alibaba's high-efficiency models, prioritizing massive knowledge capacity while maintaining low inference costs. By focusing on reasoning and agentic capabilities, Hy3 preview provides a cost-effective alternative for developers building autonomous systems that require high-performance logic without the overhead of traditional monolithic models.

You can deploy the model for complex agentic workflows and multi-step reasoning tasks that require high cost-efficiency. The weights are available on Hugging Face for self-hosting, and the model is live on OpenRouter, matching OpenCode's adoption of long-context models with a free usage tier available through May 8th, 2026.

View the full update on hy.tencent.com
Tencent Hy
Tencent Hy
@TencentHunyuan
X

👋Hi /haɪ/, we're the Tencent Hy /haɪ/ team🐧 Today, we open source Hy3 preview (295B A21B), a leading reasoning and agent model in its size, with great cost efficiency. Give us feedback to help improve Hy3 official! 🤗 https://t.co/WNuWahHk3j 📖 https://t.co/oGi2WAcreu https://t.co/GcIqb5N6MS

44retweets544likes
View on X

Still wondering? A few quick answers below.

Hy3 preview is a large language model from Tencent designed for reasoning and agentic tasks. It uses a Mixture-of-Experts architecture, meaning it has a large total capacity of 295 billion parameters but only uses 21 billion parameters per request. This design allows the model to handle complex logic while remaining efficient to run.

Yes, Tencent has open-sourced the weights for the Hy3 preview model. Developers can download and host the model themselves by accessing the repository on Hugging Face. This open-weight release allows for private deployment and fine-tuning for specific enterprise or research use cases without relying on a proprietary API provider.

You can access Hy3 preview by downloading the weights from Hugging Face for local or cloud hosting. Alternatively, the model is available through the OpenRouter API. Tencent is offering a free usage tier on OpenRouter until May 8th, 2026, allowing developers to test its reasoning and coding capabilities before committing to a paid plan.

Hy3 preview supports a context window of 256,000 tokens. This large window allows the model to process and reason across extensive documents, entire codebases, or long conversation histories in a single interaction. This capability is particularly useful for building retrieval-augmented generation systems and autonomous agents that need to maintain deep context over multiple steps.

Every HeadsUpAI update is written based on its original source and reviewed before it's published. Read our editorial standards →

Share this update