Wan 2.2 14B Shard 6 -- DMX Compressed
Original: 7,244 MB (FP32) | DMX: 1,488 MB | Savings: 79.5%
DMX-compressed shard 6 of the Wan 2.2 I2V 14B model using BFP compression. This is a single shard demo -- full model compression coming soon.
Verification
- 142/142 tensors PASS (cosine similarity > 0.99)
- Worst cosine similarity: 0.9988
- Full report: verification_report.json
How to Decompress
git clone https://github.com/willjriley/dmx
python dmx/dmx_cli.py decompress wan22_shard6.dmx shard6.safetensors
About DMX
DMX achieves 60-80% compression on neural network weights. Patent Pending. https://github.com/willjriley/dmx
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support