Unweight: how we compressed an LLM 22% without sacrificing quality

Cloudflare introduced Unweight, a lossless LLM weight-compression system that cuts model size by 15% to 22%.

On Llama-3.1-8B, it saves about 3 GB of VRAM while preserving outputs.

The system uses on-chip decompression and workload-specific execution strategies on NVIDIA H100 GPUs.

By shrinking weights without changing quality, Cloudflare can serve more models per GPU, lowering costs and improving inference speed.