# Tencent Hunyuan Open-Sources Hy4 Quantized Compression Model 'Hy4 Preview Lite'

- Category: Flash (快讯)
- Published: 2026-09-02T11:48:00.000Z
- Source date: 2026-09-02
- HTML: https://www.88z.ai/en/news/flash-20260902-e6db86bca19ff94c
- Markdown: https://www.88z.ai/en/news/flash-20260902-e6db86bca19ff94c.md
- Other language: https://www.88z.ai/news/flash-20260902-e6db86bca19ff94c.md

Tencent Hunyuan has released the Hy4 preview lite version, utilizing its self-developed Sherry sparse ternary quantization algorithm and MIX-STQ1_0 mixed-precision layer-wise quantization strategy to compress model weights from 1.5 TB to approximately 214 GB. This quantized version performs close to the original BF16 model in tasks such as long-text understanding, multi-turn context retrieval, and coding assistance.

## At a glance

- Uses Sherry algorithm and MIX-STQ1_0 strategy to compress weights from 1.5 TB to approx. 214 GB
- Performance in long-text understanding, retrieval, and coding assistance is close to the original BF16 version

## The update

Tencent Hunyuan has officially launched the Hy4 preview lite model. This version corely adopts the self-developed Sherry sparse ternary quantization algorithm, combined with the MIX-STQ1_0 mixed-precision layer-wise quantization strategy, achieving significant size reduction. Through this technical combination, the model weight data volume was drastically compressed from the original 1.5 TB to approximately 214 GB, greatly lowering storage and deployment thresholds. Despite deep compression, the Hy4 preview quantized version maintains high standards in key capabilities. In typical application scenarios such as long-text understanding, multi-turn context retrieval, and coding assistance, its actual performance is very close to the original BF16 model, ensuring that core business effects are not sacrificed while achieving lightweight deployment.

## Sources

- [腾讯混元](https://mp.weixin.qq.com/s/GS_9L54CzkgMt0vnxc2NqA) (2026-09-02)

This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.