DigiNews

Tech Watch by Johan Denoyer

← Back to articles

Lossless model compression experiment: GLM-5.2 in 25% less memory

Quality: 8/10 Relevance: 9/10

Summary

A technical bulletin reports lossless BF16 weight compression for GLM-5.2 (753B) achieving about 30.17% memory reduction with a 24.967% decoding cost. It validates across 59,509 BF16 tensors using K15 charged-format accounting and separate byte-split reconstruction, and points to open reproduction protocols and open-source artifacts.

🚀 Service construit par Johan Denoyer