Fine-tune and evaluatewithout a bill
Iterate on open weights without a cloud invoice attached to every run.
What it enables
Evaluation sweeps, quantisation experiments, LoRA fine-tunes, dataset passes — the iterative work that cloud pricing punishes — on hardware with a fixed cost of zero per run.
Reproducibility
Your model version does not change under you. Pin the weights, pin the quant, and results replicate next month.
The stack we set up for this
Iteration without an invoice attached:
- Evaluation harnesses against local endpoints — runs cost electricity
- Fine-tune and LoRA tooling on open weights, pinned and reproducible
- Whisper + Parakeet for interview and lecture transcription, offline
Reproducibility as a feature
Pin the weights, pin the quant, and results replicate next month — no silent model deprecations, no provider-side changes under your experiment.