Cloudflare open sources a lossless LLM compression tool that cuts costs, preserves accuracy, and enables decentralised AI deployment at scale.
Cloudflare has open-sourced Project Pipit, releasing a lossless compression tool that could fundamentally reshape how large language models are distributed and deployed.
Unlike traditional approaches such as quantisation and pruning—which compromise accuracy to reduce size—Pipit compresses models without altering a single numerical value.
Outputs, probability distributions, and benchmark scores remain byte-for-byte identical after decompression, eliminating the long-standing trade-off between efficiency and fidelity.
Developed under Dr. Adaosa Okafor, Director of Machine Learning at Cloudflare, the tool uses a proprietary entropy-coding algorithm to achieve fully reversible compression. Benchmarks show compression ratios of approximately 5.2× on dense Llama-3 class models and 3.8× on Mixture-of-Experts architectures.
Integrated with Cloudflare Workers AI, Pipit enables models to be stored in compressed form, streamed from edge locations, and decompressed with near-zero latency. For models exceeding 70 billion parameters, reduced network transfer times outweigh the minimal runtime overhead.
The tool is designed for immediate adoption, supporting PyTorch and SafeTensor formats without requiring changes to existing pipelines.
Beyond technical gains, the open-source release signals a strategic shift. By reducing storage and bandwidth costs by up to five times, Pipit weakens dependence on centralised GPU cloud providers and enables viable edge, on-premise, and even consumer-grade deployments.
If widely adopted, Pipit could emerge as a standard compression layer for AI models, accelerating decentralised AI and redefining global model distribution.
