Rent intelligence today. Own it tomorrow.
Ratchet is an OpenAI-compatible gateway. A small student model answers first, the teacher answers when the student is unsure, and every night the student learns from the teacher's answers. You end up owning the LoRA adapter and its signed training lineage, learned only from teachers whose license allows it.
- OpenAI-compatible
- Open teachers only
- Apache-2.0
The loop
You ownYou rent
The student answers first
The teacher answers the rest
Its answer is recorded
A new student is trained
01Facts you can check
Built to be checked, not taken on trust
Every tile is a fact you can check in the repository. Numbers from the simulator appear elsewhere in these docs, and each one is labelled Simulation.
02Where to go
Pick a path
Short guides first, then the reference. Every guide opens with one sentence and an everyday analogy.
Every part and how they connect: one request, confidence, the ledger, Night School and the release gate.
A small model that writes SQLite queries, trained only on teacher answers that were run on a database.
The exam on databases it never trained on: how it is scored, and what the run's results file says.
Every Nebius piece, step by step: the CPU VM, Token Factory, the GPU endpoint, the training job and the bucket.
Every HTTP route, generated from the gateway's OpenAPI schema, plus every setting and command.
Where the demo is, what to try in five minutes, and which numbers are live and which are simulation.
03Honest status
What is real today
What has run, what has not, and where the numbers come from.
- Done
Teacher calls on Nebius Token Factory built the text-to-SQL training data (22 Oct)
NVIDIA Nemotron 3 Ultra answered the questions whose SQL was checked on a database. The gateway’s own live mode and the day-1 checks have not been run yet.
- Done
A text-to-SQL adapter, trained on a free Colab T4
A LoRA on Nemotron 3 Nano 4B, trained from teacher answers that were run on a database. See the results.
- Not yet
The student on a Nebius GPU
The Serverless Endpoint and the nightly training job are written and tested against fakes. They have not run on a Nebius GPU yet.
- Simulation
Day-by-day savings and releases
These curves come from mock models in the simulator, and every such number is labelled Simulation.
Status as of Oct 22, 2026. The For judges guide keeps the full list.
04How it works
One loop, three speeds
The same gateway runs three loops at different speeds. The student and the teacher wear the same colours on every page.
Every request
The student samples its answer three times. If the samples agree and the confidence reaches the version’s threshold, its answer is served; if not, the teacher answers. A few served answers are audited by the teacher.
Every few minutes
The gate reads the audits and moves a new version forward (shadow, then 10%, 50% and 100% of traffic) or rolls it back.
Every night
A fresh adapter is trained on the teacher’s answers only, tested on a held-out golden set, and handed to the gate.
05Why it matters
Cheaper, yours, license-clean and safe to release
Cost
Pay the teacher only when the student is unsure. Savings are reported per token and by GPU-hours paid, and the report says when an endpoint costs more than it saves.
Read more: CostOwnership
You keep the LoRA adapter and its signed training lineage. The Nano base model keeps NVIDIA’s license.
Read more: OwnershipLicense-clean
Only teachers whose license allows training are accepted. Each stored example records its teacher and license.
Read more: License-cleanSafe releases
A new adapter passes a held-out exam, a shadow run and a 10%, 50%, 100% canary, with automatic rollback.
Read more: Safe releases
06Built with
Open models and open tools
Why "Ratchet"
A ratchet turns only one way.
Every night more of your traffic moves from the model you rent to the adapter you own, and the gates are the pawl that stops it from slipping back: a bad release is rolled back automatically.