Compressed models, measured honestly.
Exeaon takes a trained model and rewrites its weights into a much smaller form — no retraining, no access to the original training data, no change to the model architecture. What follows is what we measured, including the parts that are not finished.
Dense transformer, 14B
AvailableScored fractionally above the original. That is measurement noise, not a gain — but it means the loss is below what this benchmark can resolve.
Dense transformer, 8B
AvailableLarger models absorb compression better; the loss shrinks as the model grows.
Dense transformer, 4B
AvailableReference configuration. Runs on a single consumer GPU.
Mixture-of-experts, 30B-A3B
AvailableSparse experts compress as well as dense layers, not worse.
Dense transformer, 0.6B
AvailableSmall enough to serve from CPU. Published as Exeaon1-Dzo-0.6B.
Video diffusion transformer
Available98.2% of the transformer compressed; output verified against the original frame-by-frame at 26.3 dB PSNR.
Speech recognition, 809M
AvailableWord error rate is unchanged at 6.19%, and transcriptions are character-identical to the original on the evaluation sample. Compression changed nothing measurable.
Vision-language, 2B
ResearchBelow our release bar on text-only calibration. Held back until multimodal calibration closes the gap.
A number is only useful if you can check it.
Accuracy retained is the mean score across four standard reasoning benchmarks — ARC-Easy, ARC-Challenge, HellaSwag and PIQA — as a percentage of the uncompressed model's score.
Both models are scored on the same GPU, with the same evaluation harness at the same version, over the same number of samples. Changing any one of those moves the result on its own, which is how compression gets credited for wins it did not earn.
For the video model there is no accuracy benchmark, so we generate from the original and the compressed model with an identical prompt, seed and step count, and compare the frames directly.
The runtime is a public Python package under Apache-2.0. It loads a compressed model and serves it like any other — on GPU, or on CPU where the model is small enough.
pip install epure-runtime
import epure
model, tok = epure.load("model.ebin")The runtime is free and stays free. The compression method that produces these models is proprietary to Zenux Plimver Technologies.
Interested in compressing your own models?
We work with teams serving large models in production, where footprint and memory are the binding constraint.
Get in touch