dinov2-large-uc-v3-bpw5

DINOv2-Large -- UltraCompress v3 (5-bit, near-lossless)

UltraCompress is retired. Its public materials preserve historical research; new API access and paid pilots are no longer offered. Results below apply to their recorded evaluation conditions.

The public v0.6.27 CLI checks basic pack structure and computes SHA-256 fingerprints unless hashing is skipped. It does not reconstruct weights, reproduce the evaluations below, or automatically compare its fingerprint with a trusted reference.

First public near-lossless 5-bit Vision Transformer. Compressed from facebook/dinov2-large using the UltraCompress v3 codec -- the same pipeline that ships 22 transformer LLM packs at a mean PPL ratio of 1.005x.

This pack proves the codec generalizes beyond text transformers to vision architectures with zero format changes.

Quality

CLS-token cosine is the natural quality metric for DINOv2: it measures how closely the compressed model's learned representation matches the original across a diverse image set. 0.9988 means the compressed features are effectively interchangeable with the originals for downstream tasks (kNN classification, retrieval, linear probing).

Compression Details

Comparator Note

For vision transformers, CLS-token cosine replaces perplexity (PPL) as the quality comparator. PPL is undefined for non-autoregressive models. The CLS-token cosine of 0.9988 on DINOv2-Large is comparable in quality preservation to the 1.004x PPL ratio achieved on similarly-sized LLMs (e.g., Qwen3-8B at 1.0044x).

Public CLI checks

pip install "ultracompress==0.6.27"
hf download SipsaLabs/dinov2-large-uc-v3-bpw5 --local-dir ./pack
uc verify ./pack

The public v0.6.27 CLI checks basic pack structure and computes SHA-256 fingerprints unless hashing is skipped. It does not reconstruct weights, reproduce the evaluations below, or automatically compare its fingerprint with a trusted reference.

Pack Format

UltraCompress v3 binary (.uc files). See the public repository for historical documentation; the public CLI contains no codec implementation. This pack uses the same v3 format as all LLM packs -- no format extensions were needed for vision transformers.

License

Apache 2.0 (same as base model).

Citation

@misc{ultracompress2026,
  title={UltraCompress: Near-Lossless 5-bit Compression for Large Neural Networks},
  author={Sipsa Labs},
  year={2026},
  url={https://github.com/sipsalabs/ultracompress}
}