SipsaLabs/qwen3-8b-uc-v3-bpw3
SipsaLabs/qwen3-8b-uc-v3-bpw3 is an 8 billion parameter language model, a 3-bit compressed version of Qwen/Qwen3-8B, developed by Sipsa Labs using their UltraCompress technology. This model offers a significantly reduced size while maintaining a cryptographically verifiable and independently perplexity-verified reconstruction. It is designed for efficient deployment where model size is a critical factor, providing a balance between performance and resource consumption.
Loading preview...
Model Overview
SipsaLabs/qwen3-8b-uc-v3-bpw3 is an 8 billion parameter language model, specifically a 3-bit compressed variant of the original Qwen/Qwen3-8B model. This compression was performed by Sipsa Labs using their proprietary UltraCompress technology.
Key Characteristics
- 3-bit Compression: This model utilizes a 3-bit compression scheme, which is a lossy process resulting in a larger perplexity drift compared to 5-bit compression methods.
- Reproducible Reconstruction: It features a reproducible and cryptographically verifiable reconstruction process, ensuring a deterministic decode to a SHA-256-pinned validated artifact. It's important to note that this reconstructed model is not bit-identical to the original bf16 model.
- Perplexity Verified: The model has undergone independent end-to-end perplexity verification, confirming its performance post-compression.
- License: The model operates under the BUSL-1.1 license, which includes an Additional Use Grant.
Usage and Limitations
This model is suitable for applications where a significantly reduced model footprint is critical, and a slight trade-off in perplexity due to 3-bit compression is acceptable. The UltraCompress technology, while retired for public API access and paid pilots, preserves historical research. The public v0.6.27 CLI for UltraCompress can check pack structure and compute SHA-256 fingerprints, but it does not reconstruct weights or reproduce the evaluations mentioned. For more details, refer to the Sipsa Labs documentation and their GitHub repository.