Pilots
How a pilot works.
Scope, compression, evaluation, and delivery for a Minutus model-compression pilot.
01
Scope
Tell us the model (open or your own fine-tune), the target hardware and memory budget, and the tasks that matter.
02
Compress
We compress the model, on our infrastructure or yours. Typical turnaround: 1-2 weeks.
03
Measure
You receive a quality report on standard benchmarks and on samples of your own tasks, side by side with the uncompressed model and standard 4-bit / 8-bit baselines.
04
Decide
Keep the compressed model under a pilot license and discuss production deployment.
What we need from you
- Model access (weights or Hugging Face ID)
- Target memory budget and hardware
- Optional: a sample of representative prompts
What you get
- Compressed model as Hugging Face checkpoints
- Evaluation report
- Recommended bit width for your use case
Your data stays yours
We use only a small sample of text to tune compression, and it can stay on your premises.
Privacy Policy →Supported today
Dense transformer language models from about 1 to 8 billion parameters, e.g. Qwen3 and Apertus, including fine-tuned versions. Larger models are on the roadmap.
Output format: PyTorch / Hugging Face checkpoints; optimized runtime formats in development.