SVD-Surgeon: Optimal Singular-Value Surgery for Large Language Model Compression
Paper • 2606.23568 • Published
Qwen3.5-0.8B-SVD-Ob-1.54x
SVD-Surgeon OBS-Compensated Compression (arXiv:2606.23568)
Compressed with weightnoise --method svd-observe at keep_ratio=0.5.
Benchmark Results
| Benchmark | Baseline | SVD-Observe (1.54x) | Delta |
|---|---|---|---|
| Short Text PPL | 45.50 | 45.50 | -0.01 |
| Fibonacci JS | — | — | YES |
| Logic Puzzle | — | — | YES |
| Word Reverse | — | — | YES |
Verdict
PASS: Compression preserves model quality within 0.0002 loss delta.
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("KiriLabs/Qwen3.5-0.8B-SVD-Ob-1.59x")
tokenizer = AutoTokenizer.from_pretrained("KiriLabs/Qwen3.5-0.8B-SVD-Ob-1.59x")
- Downloads last month
- 9
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support