Exeaon1-Kese-32B

Exeaon1-Kese-32B

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

Requires epure-runtime>=0.2.4. Earlier versions load every model as float16. Qwen3-32B is trained in bfloat16 and its activations exceed the float16 range, so on 0.2.3 this model overflows to infinity partway through the second layer and produces garbage on prompts of ordinary length. This is not a soft warning: the failure is silent on short prompts and total on real ones.

pip install "epure-runtime>=0.2.4"

Quality

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

Retention: 100.2% of base mean accuracy.

The compressed model scores fractionally above the original. That is not a gain — it is measurement noise at limit=300, where one item is worth 0.33 points. What it does mean is that the loss from compression is smaller than this benchmark can resolve.

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

Usage

pip install "epure-runtime>=0.2.4"

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]))

Notes on speed

On a large GPU this runs slower than dense fp16. A vendor tensor-core GEMM is heavily tuned and a large card 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.

Qwen3-32B needs roughly 61 GB in bfloat16. This container needs 16.42 GB on disk and stays packed in memory.

Licence

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