Exeaon1-Kese-30B-A3B

Compressed with E-PURE. Runs with the free epure-runtime, and stays compressed in memory -- the dense weight is never assembled.

Base model	Qwen/Qwen3-30B-A3B
Size on disk	16.27 GB
Compression	3.50x
Format	.ebin

Quality

Measured against the base model on the same GPU, the same harness version and the same limit. Not copied from anyone's README.

benchmark	base	Exeaon1-Kese-30B-A3B	delta
ARC-Challenge	53.67	50.33	-3.34
ARC-Easy	77.67	75.67	-2.00
HellaSwag	64.67	63.67	-1.00
PIQA	81.33	82.33	+1.00
mean	69.34	68.00	-1.34

Retention: 98.1% of base mean accuracy.

A mixture-of-experts model, where most parameters sit in expert layers that any one token never touches. Sparse experts compress as well as dense layers here, not worse -- which is the result that was not obvious going in.

Scored at limit=300, so roughly 2-3 points of standard error per task. Read the mean, not any single row.

Speed and footprint

	base	Exeaon1-Kese-30B-A3B
decode, batch 1	35.1 tok/s	2.5 tok/s
decode, batch 8	147.7 tok/s	17.9 tok/s
peak VRAM	58.06 GB	32.16 GB

Read this honestly: on GPU we are slower than dense fp16. A vendor tensor-core GEMM is heavily tuned and a large GPU has bandwidth to spare, so trading compute for memory loses there. The win is fitting in less memory -- running where the dense model does not fit at all, on a smaller card, or alongside more of them.

Measured on NVIDIA RTX PRO 6000 Blackwell Server Edition.

Usage

pip install epure-runtime

import epure

model, tok = epure.load("model.ebin", device="cuda")
print(tok.decode(model.generate(**tok("Hello", return_tensors="pt").to("cuda"),
                                max_new_tokens=50)[0]))

Licence

The runtime is Apache-2.0 and free. The compression method that produced this container is proprietary to Zenux Plimver Technologies LTD.