---
title: "AngelSlim Shrinks Tencent's 1.5TB Hy4-Preview to 214 GiB at 6x Compression"
subtitle: "Tencent's AngelSlim team shrinks the 770B Hy4-preview MoE from 1.5TB down to 213GB using a custom 1.31-bit quantization strategy with minimal accuracy loss."
url: "https://alphasignal.ai/news/angelslim-shrinks-tencent-s-1-5tb-hy4-preview-to-214-gib-at-6x-compression"
author: "AlphaSignal Newsroom"
author_url: "https://alphasignal.ai/editorial-team"
company: "AlphaSignal"
company_url: "https://alphasignal.ai/company/alphasignal"
published: "2026-08-28T14:40:23.000Z"
updated: "2026-09-05T07:42:52.495Z"
type: "Model"
topics:
  - "Llms"
  - "Gpus"
subtopics:
  - "Small Models"
  - "Quantization"
upvotes: 44
source_url: "https://huggingface.co/AngelSlim/Hy4-preview-GGUF"
comments: 0
truncated: true
---

# AngelSlim Shrinks Tencent's 1.5TB Hy4-Preview to 214 GiB at 6x Compression

> Tencent's AngelSlim team shrinks the 770B Hy4-preview MoE from 1.5TB down to 213GB using a custom 1.31-bit quantization strategy with minimal accuracy loss.

Tencent's Hy4-preview is a 770B parameter Mixture-of-Experts model. In its native form it weighs 1.5TB and requires a rack of B200s to run. The [AngelSlim team](https://huggingface.co/AngelSlim/Hy4-preview-GGUF) just released three GGUF builds that bring it within reach of a single well-specced node, including one that compresses the whole thing to roughly 213 GiB while keeping benchmark scores within a couple of points of the BF16 original.

That compression, from 1.5TB to about 214 GiB, relies on a new mixed-precision recipe called MIX-STQ1\_0\. Some layers drop to 1.31 bits per weight while higher-sensitivity layers keep more bits, with the allocation driven by an importance matrix rather than uniform rounding.

## Three builds, three tradeoffs

The repo ships three quantizations of the same [Hy4-preview](https://huggingface.co/tencent/Hy4-preview) checkpoint:

| File      | Size    | Bits/weight | Use case                            |
| --------- | ------- | ----------- | ----------------------------------- |
| Q4\_K\_M  | 435 GiB | 4.86        | Safe default when VRAM is available |
| UD-IQ1\_M | 220 GiB | 2.44        | Unsloth's dynamic 1-bit strategy    |
| STQ1\_0   | 214 GiB | 2.38        | MIX-STQ1\_0 recipe, smallest        |

None of these run on stock llama.cpp because the hyv4 architecture is not yet upstream. You need to apply patches from the hy4-preview-patch directory, after which standard commands like `llama-cli` and `llama-bench` work as usual. On an 8x H20 rig, the STQ1\_0 build hits roughly 204 tokens/sec prefill and 20 tokens/sec decode.

## How STQ1\_0 works

STQ1\_0 originates from llama.cpp PR #22836\. Each weight takes one of three values (negative d, zero, or positive d), and exactly one lane out of every four is forced to zero, producing 3:4 sparsity. Each group of 4 weights is stored as a 4-bit code plus a 1-bit table-select that indexes into a 32-entry codebook, with a single fp16 scale shared across 256 weights. That works out to 42 bytes per 256 weights, or 1.3125 bits per weight.

_Preview only. Full text requires an AlphaSignal Pro account._

## Sources

- [AngelSlim/Hy4-preview-GGUF](https://huggingface.co/AngelSlim/Hy4-preview-GGUF)
- [Tencent Hunyuan on X](https://x.com/TencentHunyuan/status/2093572224342954019)
- [tencent/Hy4-preview](https://huggingface.co/tencent/Hy4-preview)
- [Tencent-Hunyuan/Hy4-preview GitHub](https://github.com/Tencent-Hunyuan/Hy4-preview)
- [What Is Hunyuan 4? Hy4 Preview Pricing & Benchmarks](https://gptproto.com/news/what-is-hunyuan-4)