Privacy
Updated 28 September 2026
We do not train on your data.
Configuration and performance metadata
quantization, engine flags, placement, tokens per second, latency, cost
kept and used to plan future jobs, yours and other people's
Your weights, eval data and request content
prompts, completions, uploaded datasets, model weights
never used for training, deleted on request, removed from every provider we placed them on
That split is the whole policy. We benefit from learning that fp8 on an H100 runs a Qwen2 at 127 tokens a second. We do not need to read your prompts to learn it.
Where your data goes
Your model comes from Hugging Face and your eval from the dataset you name or the file you upload. Both are read onto the GPU host that runs your job.
The record of a search, its measurements and the requests used to take them, is archived in Cloudflare R2 so the search can resume on another host if one fails. Quantized weights we produce for you are kept there for seven days, then expire.
Hosts we rent
When none of our own hardware fits your job, we rent a GPU host for it from a marketplace, vast.ai today. The key to our marketplace account never goes onto that host. It is destroyed at the provider, disk included, ten minutes after your job lets it go.
What we keep about you
Your name, email address and Google account id, from signing in with Google. We store the address of your profile picture and never display it.
Deleting
Deleting a job removes its weights, eval data and outputs from our storage and from every provider inside 24 hours, and writes a deletion record you can request.
Who else handles it
- signing in
- Hugging Face
- model and dataset downloads
- Cloudflare
- R2 storage for search records and weights
- vast.ai
- GPU hosts we rent for a job
- Railway
- hosting this site
Questions, or a deletion request: write to us.