Breaking

Tencent compresses Hy4 preview from 1.56TB to 213.66GB with mixed-precision quantization

September 1, 2026 at 20:47 EDT

Foundation Models
Open Source

Tencent has released a mixed-precision GGUF of its open-weight Hy4 preview model that cuts the checkpoint from about 1.56TB to 213.66 GiB—an 85% size reduction—while reporting only small drops on several internal benchmarks. The extreme build, Hy4-preview-STQ1_0.gguf, is meant to make the 770-billion-parameter Mixture-of-Experts system easier to store and serve without matching the original BF16 footprint.

September 2026 · Tencent Hunyuan

Hy4 preview, 85% smaller

A mixed-precision GGUF compresses the open-weight MoE from about 1.56TB to 213.66 GiB, keeping most reported scores close to BF16.

85%

size cut vs ~1.56TB

213.66

GiB STQ1_0 GGUF

2.38

average bpw

~7×

smaller than BF16

Footprint at scale

One block ≈ one-seventh of the original size. Seven blocks (BF16) versus one block (STQ1_0).

~1.56TB

BF16 baseline

213.66 GiB

MIX-STQ1_0

Takeaway: about seven times smaller on disk, not a uniform 1-bit squeeze.

How bits are assigned

1. Calibrate layers

→

2. Protect sensitive tensors

→

3. Squeeze expert bulk to ~1.31-bit

Held up

MCP Atlas 83.7→83.2 and MRCR 81.3→81.1. Long-context work stayed close to BF16 on Tencent’s tests.

Still hard

SWE-Bench multi fell 1.6 points. Files need hyv4 llama.cpp patches, and ~214 GiB still will not fit a typical workstation.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.