# Tencent Hunyuan highlights Hy4 preview compressed to 214 GiB

Tencent Hunyuan quoted a Zhihu Frontier explainer on shrinking Hy4 preview's 770B-parameter weights from about 1.5TB to 214 GiB with mixed-precision quantization. The quoted write-up says the parameter count is unchanged and reports nearly unchanged MRCR with a small decline in math.

Language: en
Time zone: America/Los_Angeles
Current time (UTC, cached as of): 2026-11-10T07:31:00.000Z
Query date (America/Los_Angeles): 2026-11-09

HTML: https://didcodexreset.com/news/023f8b753f514cfaa3899c5e.html

Content language: en
Localization state: sameLanguage

Source: Tencent Hy · Published 9/21/2026, 19:46:31

Tencent Hunyuan ([@TencentHunyuan](https://x.com/TencentHunyuan)) quoted Zhihu Frontier with the comment "Less size, same intelligence," pointing to an explainer on compressing **Hy4 preview**.

In the quoted post, Zhihu Frontier ([@ZhihuFrontier](https://x.com/ZhihuFrontier)) said contributor yghstill of Tencent Hunyuan's quantization team described packing the **770B**-parameter model by shrinking weights from roughly **1.5TB to 214 GiB**. The parameter count remains 770B; the compression changes how those weights are represented. The explainer names the Sherry quantization algorithm, STQ1_0 storage format, and MIX-STQ1_0 mixed-precision allocation, and says the complete mixed-precision model averages about **2.38 bits per weight**, with aggressive compression concentrated on expert weights.

The quoted account says the team implemented STQ1_0 CUDA kernels in a patched llama.cpp build and reports nearly unchanged MRCR retrieval performance with a small decline in math. Against UD-IQ1_M at a similar bit budget, it says the mixed-precision model led on the reported evaluations, including a gain of more than five points on MRCR. Tencent Hunyuan's own post does not provide those measurements.

Tags: Hunyuan, Hy4, Quantization, Tencent

[View original post](https://x.com/TencentHunyuan/status/2102225827903807939)