Open-weight development

Help train the next open-weight model

Hosted API revenue funds open-weight releases. This campaign accelerates training and lets experiments run continuously instead of waiting on cloud budgets.

  • 3 releases shipped
  • $540 saved per training run
  • RTX 4090 target
  • Results published publicly
github.com/withoutbg/withoutbg
NVIDIA RTX 4090 Founders Edition, target hardware for the GPU fund

GPU fund

Goal: 1 x NVIDIA RTX 4090

Estimated impact

  • faster iteration
  • reproducible experiments
  • more open releases

Why fund this?

API supports open weights

Hosted API revenue pays for production infrastructure and funds open-weight releases when model work transfers.

Local GPU removes rerun costs

The last full AWS run cost $540. Owned hardware makes failed runs and ablations cheaper to repeat.

Faster experimentation

Continuous access means more validation runs, tighter feedback loops, and fewer pauses between promising changes.

Commitment

What happens after retraining

If the retraining improves validation metrics, the model will be released publicly. If not, findings will still be published.

Donations here go toward open-weight model R&D hardware only. API infrastructure is paid for by API customers.

Training details are documented in Engineering Background Removal. More context on the solo-founder setup is in About.

Funding options

Contribute to the training fund

Pick a one-time amount or set a small monthly commitment. Stripe handles payment; no account is needed.