SipsaLabs/qwen3-1.7b-uc-v3-bpw3
SipsaLabs/qwen3-1.7b-uc-v3-bpw3 is a 2 billion parameter language model, a 3-bit compressed version of Qwen/Qwen3-1.7B, developed by Sipsa Labs using UltraCompress. This model offers a lossy compression with a 32768 token context length, providing a cryptographically verifiable and independently perplexity-verified reconstruction. It is designed for efficient deployment where reduced model size and verifiable integrity are critical.
Loading preview...
Model Overview
SipsaLabs/qwen3-1.7b-uc-v3-bpw3 is a 2 billion parameter language model derived from Qwen/Qwen3-1.7B, developed by Sipsa Labs using their UltraCompress technology. This version applies a 3-bit lossy compression, resulting in a smaller footprint suitable for resource-constrained environments. It maintains a substantial context length of 32768 tokens.
Key Characteristics
- 3-bit Compression: Utilizes a 3-bit lossy compression scheme, leading to a smaller model size compared to its original bf16 counterpart. Note that this compression introduces a larger perplexity drift than 5-bit compression lines.
- Verifiable Reconstruction: Features a reproducible and cryptographically verifiable reconstruction process, ensuring a deterministic decode to a SHA-256-pinned validated artifact. This artifact is not bit-identical to the original bf16 model.
- Perplexity Verified: The model's end-to-end perplexity has been independently verified, providing a measure of its language modeling quality post-compression.
- License: Distributed under the BUSL-1.1 license with an Additional Use Grant.
Usage and Verification
Developers can interact with the model using the ultracompress Python package. While the public v0.6.27 CLI checks basic pack structure and computes SHA-256 fingerprints, it does not reconstruct weights or reproduce the full evaluations. Further documentation and verification details are available on the Sipsa Labs website and their GitHub repository.