FAQ

Questions about model compression.

Answers about supported models, compression size, evaluation, data handling, and pilot delivery.

Which models can you compress?

Dense transformer language models from about 1 to 8 billion parameters today, including Qwen3 and Apertus, and fine-tuned versions of them. Larger models are on our roadmap.

How small do models get?

Typically 3 bits per weight: Qwen3-8B goes from 16.4 GB to 3.90 GB, Qwen3-4B from 8.04 GB to 2.45 GB. 2 bits is a research preview.

Do you retrain the model?

No. Compression uses a small text sample and does not require a training dataset.

How do you measure quality?

On standard task benchmarks (instruction following, coding, knowledge) and on samples of your own tasks, always side by side with the uncompressed model and standard 4-bit / 8-bit baselines.

Does my data leave my company?

No. Compression can run on your infrastructure, and only a small text sample is needed.

What do I receive?

The compressed model as Hugging Face checkpoints plus an evaluation report, typically within 1-2 weeks.