Research / 02

Work that can
be inspected.

Research is useful when someone else can challenge it, rerun it and build on it. We publish the method, the artifacts and the uncomfortable details.

Open arXiv:2603.04162
BS / RESEARCH02

Bielik Q2 Sharp

How small can a model get
without losing its language?

The published study, led by BitSharp founder Jakub Prejzner, compares six 2 bit quantization methods across 22 Polish benchmark suites. It measures the tradeoff instead of hiding it.

FP16 baseline22GB
Quantized model3.26GB
6.7× compression71.92% benchmark aggregate22 benchmark suites

Public infrastructure

The result is not
a screenshot.

The paper is only the index. Models, datasets, Hessians and evaluation traces make the work inspectable.

01
Models9 public releases
02
Datasets3 public datasets
03
PublicationarXiv:2603.04162
04
Agent evaluationPolAgentBench, 67 task draft
In development

Method

01

Make the claim precise.

No broad quality label. Name the metric, baseline and evaluation set.

02

Expose the route to it.

Keep the code, weights and evaluation artifacts close to the conclusion.

03

State what did not work.

A real result includes caveats, failure modes and methodological limits.

Build from evidence

Research is the engine.
Production is the test.

See the system Work with BitSharp