Hy3 is Tencent’s open large language model built for reasoning, coding, long-context work, and AI agents. This guide explains what Hy3 is, how it is released, and where it fits for developers comparing open models.
What Is Hy3?
Hy3 is a large language model from Tencent’s Hy Team. It uses a Mixture-of-Experts architecture, often shortened to MoE.
The model has 295B total parameters, but it activates 21B parameters per token. That design helps Tencent target stronger output quality without running the full model on each token.
Hy3 follows the earlier Hy3 Preview release. Tencent says the full Hy3 release used feedback from more than 50 product teams and a larger post-training pipeline. The model focuses on reasoning, coding, tool use, long-context tasks, and agent workflows.
Hy3 is not a 3D generation model. It is a text-generation model for chat, coding, tool calling, and structured work.
Overview
| Item | Details |
|---|---|
| Model name | Hy3 |
| Developer | Tencent Hy Team |
| Model type | Open large language model |
| Architecture | Mixture-of-Experts |
| Total parameters | 295B |
| Active parameters | 21B |
| Context length | 256K tokens |
| License | Apache 2.0 |
| Open weights | Yes |
| Main formats | BF16 and FP8 version |
| Best fit | Agents, coding, reasoning, long-context tasks |
| Deployment | vLLM, SGLang, Transformers |
Features
MoE Architecture
Hy3 uses 192 experts and activates the top 8 experts during inference. This setup gives the model a large total capacity while keeping active compute lower than a dense model of the same total size.
256K Context Window
Hy3 supports a 256K token context length. You can use it for long documents, multi-file code review, knowledge-base Q&A, and workflows that need long conversation memory.
Agent and Tool Calling Support
Tencent positions Hy3 as an agent-focused model. The model card lists support for tool calling, output constraints, and multi-turn task handling.
This makes Hy3 useful for coding agents, office agents, file workflows, and apps that need structured API calls.
Coding and Productivity Tasks
Hy3 targets coding, frontend development, financial modeling, office work, and game development tasks. Tencent reports blind expert evaluations across real work tasks, not only public benchmarks.
Use your own tests before replacing a production model. Hy3 looks strongest when the task needs planning, tools, and long context.
Better Multi-Turn Tracking
Tencent says Hy3 improves coreference resolution, ellipsis recovery, and constraint inheritance across long conversations. In practice, that means the model should keep track of earlier instructions better during agent sessions.
Who Should Use Hy3?
Hy3 makes the most sense for developers building agent systems, coding assistants, long-context chat apps, and enterprise workflows.
It may not fit small local setups. The model files are large, and practical serving needs serious GPU memory. For lightweight chatbots, a smaller model will be easier to run.