TironianThe plugin is free. Training costs cents.
Tironian itself costs nothing and never will. The only cost is GPU time when you train a model, and you choose where that comes from.
Capture, grading, task families, the report, the Desktop page, routing and fallback.
- MIT licensed
- No account, no telemetry
- Nothing leaves your machine
A GPU in your own Modal account, billed by the second. Modal's monthly free credit usually covers it.
- Cost shown before every run
- Hard cap, enforced as a timeout
- Files deleted from Modal afterwards
Any NVIDIA card with 8 GB or more. Same training script, nothing leaves your machine.
- One-time download of the training libraries
- Your electricity, nobody's cloud
What would it cost you?
One run
Per month
About 0 min of GPU time per run.
Estimate from Modal's per-second GPU prices and Tironian's own timing model, including container start and GGUF export. Your Modal dashboard shows the real number.
The other side of the ledger.
Every request a small model answers is one your paid model doesn't. Tironian counts it: each served request is credited with the average cost of that task on your own model, and the total shows on the Desktop page and in /tironian. If your model is free, the saving is time and privacy rather than money, and Tironian shows that honestly as $0.
Do I need a card for Modal?
No. Modal's Starter plan gives a small free credit without a card, enough for a couple of runs. Adding a card raises the monthly free credit; you are only charged above it, and you can set a spending limit in Modal.
Can a run cost more than the estimate?
Not more than your cap. max_usd_per_run becomes the GPU job's timeout, so the job is stopped before it can cost more. The default cap is $1.00.
What does running the model cost?
Nothing extra. The trained model runs on your own computer through Hermes's local model runtime.