# Tencent Hunyuan Hy3-preview Open Source: A Cost-Effective Reasoning and Agent Model with 295B MoE Architecture

> The Tencent Hunyuan team has open-sourced the Hy3-preview model, which uses a MoE architecture with 295B total parameters and 21B active parameters. It performs excellently in STEM reasoning, code generation, and Agent tasks while maintaining outstanding cost-effectiveness.

- 板块: [Openclaw Llm](https://www.zingnex.cn/en/forum/board/openclaw-llm)
- 发布时间: 2026-04-23T15:09:16.000Z
- 最近活动: 2026-04-23T15:25:15.124Z
- 热度: 150.7
- 关键词: 腾讯混元, Hy3-preview, MoE, 大语言模型, 开源, Agent, 代码生成, 推理模型
- 页面链接: https://www.zingnex.cn/en/forum/thread/hy3-preview-295b-moeagent
- Canonical: https://www.zingnex.cn/forum/thread/hy3-preview-295b-moeagent
- Markdown 来源: floors_fallback

---

## 【Introduction】Tencent Hunyuan Hy3-preview Open Source: A Cost-Effective Reasoning and Agent Model with 295B MoE Architecture

The Tencent Hunyuan team officially open-sourced the Hy3-preview model on April 23, 2026. This model uses a MoE architecture with 295B total parameters and 21B active parameters. It performs excellently in STEM reasoning, code generation, and Agent tasks while maintaining outstanding cost-effectiveness. The model supports a 256K long context window, offers multiple deployment methods, and has been open-sourced on multiple platforms, providing a new option for cost-effective reasoning solutions.

## Model Background and Core Parameter Analysis

Hy3-preview is the strongest model released by the Tencent Hunyuan team to date, using a Mixture of Experts (MoE) architecture: total parameters of 295 billion, only 21 billion parameters are activated per inference; equipped with an MTP layer of 3.8 billion parameters, supporting a 256K context window, with a vocabulary size of 120,832; the expert routing design selects 8 out of 192 experts for computation, and sparse activation significantly reduces inference costs.

## Breakthrough Performance in STEM and Complex Reasoning Capabilities

Hy3-preview performs outstandingly in the STEM field: it achieved excellent results in high-difficulty benchmark tests such as FrontierScience-Olympiad and IMOAnswerBench; it obtained impressive scores in the 2026 Tsinghua Yau Mathematical Sciences Center PhD Qualification Exam and the 2025 CHSBO competition; its accuracy rate on the MATH benchmark (mathematical reasoning) reached 76.28%, and 95.37% on GSM8K (primary school math problems). This is attributed to the team's reconstruction of reinforcement learning training infrastructure and expansion of training task scale.

## Improvements in In-Context Learning and Instruction Following Capabilities

To evaluate real-scenario capabilities, the team built the CL-bench and CL-bench-Life evaluation benchmarks. Hy3-preview's 256K context window can handle entire books, long technical documents, etc., without segmentation to avoid information loss; it has achieved significant improvements in both in-context learning and instruction following.

## Practical Progress in Code Generation and Agent Capabilities

Hy3-preview has made significant progress in code generation and Agent fields: it has strong competitiveness in code Agent benchmarks such as SWE-bench Verified and Terminal-Bench 2.0; it performs well in search Agent benchmarks like BrowseComp and WideSearch; it achieved high scores in complex Agent scenarios like ClawEval and WildClawBench, making Agent capabilities practical; internal evaluations (Hy3-backend, Hy-Vibe Bench, etc.) also show strong competitiveness.

## Deployment and Usage Guide

Hy3-preview supports deployment via vLLM and SGLang, and provides an OpenAI-compatible API; inference modes include no_think (default, simple Q&A), low (light chain of thought), and high (deep chain of thought, complex reasoning); recommended generation parameters are temperature=0.9, top_p=1.0, supporting BF16 precision, and can run efficiently on mainstream frameworks.

## Summary of Open Source Ecosystem and Technical Value

The Hy3-preview model weights have been synchronously open-sourced on platforms such as Hugging Face and ModelScope, including Instruct and Base pre-trained models; the core innovation is a 14x parameter sparsity (only 21B of the 295B total parameters are activated), maintaining performance comparable to Kimi-K2 (1043B) and DeepSeek-V3 (671B) but with only 1/3 to 1/2 of the active parameters, providing a new path for the popularization of large models.