Tencent Hunyuan's AngelSlim Shrinks Hy4-preview from 1.5TB to 214GB

On September 1, Tencent's Hunyuan team released AngelSlim, compressing the flagship Hy4-preview model from about 1.5TB to 214GB — roughly a 7x size reduction and 85% weight compression — with minimal quality loss, and enabling multi-machine GPU deployment.

Confirmed

• The core is Sherry sparse ternary quantization, aggressively compressing MoE routed experts down to about 1.25 bits per weight
• It works with the MIX-STQ1.0 mixed-precision strategy, where calibration data determines per-layer compression granularity
• Per tests relayed by @QuixiAI, the MCP Atlas score only dropped from 83.7 to 83.2
• @智东西 reported the compressed model performs on par with the original on long-context understanding and multi-turn retrieval tasks
• Supports cross-machine GPU stitching and heterogeneous collaboration; local formats such as GGUF are in progress

Why it matters

• A 1.5TB model is beyond the hardware reach of most individuals and organizations; 214GB makes consumer-grade or multi-machine cluster deployment feasible
• If sparse ternary quantization proves stable on Hy4, it offers a reusable methodology for slimming other very large MoE models

2026-09-01 ~ 2026-09-02 · 5 related posts

• [source] Tencent Hunyuan AngelSlim: Compressing Hy4 Model to 214GB with Heterogeneous Inference — 腾讯混元 · 2026-09-01
• [source] Tencent Hunyuan releases lightweight model with 85% weight compression — 智东西 · 2026-09-01
• [source] Tencent's Sherry quantization shrinks Hy4 model size by 7x — QuixiAI · 2026-09-01
• Tencent Hy4-preview shrunk to 214GB via Sherry quantization — alejandroll10 · 2026-09-02

1 near-duplicate retellings: ccerrato147