OPTIC compresses per-layer embeddings 1000x with minimal performance loss
Jiacheng Liu · hf · 2026-10-27
OPTIC, shared on Hugging Face, achieves extreme compression of per-layer embeddings using token-shared indexed codes plus layer-specific affine decoders fitted to the output path, keeping the backbone frozen.
Benchmark charts show 85% performance retention at roughly 1000x resident-byte compression, while PQ and OPQ-P baselines collapse to near zero between 100x and 500x, with minimal inference overhead.
More from Infra
- Running the π0.5 VLA model locally on a 16GB RTX 4070 Ti SUPER with APXInf — lxfater · 2026-10-27
- MLOps vs LLMOps vs AIOps: One Chart Maps the 5 Key Stages — _jaydeepkarale · 2026-10-27
- GPS spoofing highway thieves steal millions in data center-bound cargo — luisdans · 2026-10-27
- The 'GPU-poor' hack: befriend top-university professors and casually mention you lack compute — vxnuaj · 2026-10-27
- Cadence launches Maestro AI agents for chip design, claiming up to 30% quality gains — sven_ai · 2026-10-27
- Prediction: AI labs must find new high-value workloads to justify infra spend — yi111 · 2026-10-27