Model versions and publishing

SDK examples on this page require the 0.5.0 source preview. npm publication is pending; use the HTTP endpoints directly in the meantime.

A model family is your application's stable model identity. Each successful training run produces an immutable version. Your application decides which version to publish.

All three resources belong to the organization that created them. Knowing an ID does not grant another organization access. A model family without a published ready version cannot serve inference.

Create a model family

const model = await rebyte.models.create({
  name: 'Order extraction',
  base_model: 'Qwen/Qwen3.6-35B-A3B',
  scenario_id: scenario.id,
}, { headers: { 'Idempotency-Key': 'orders-model-001' } });

Use the family's Scenario to review and select training data. A training run saves the starting source and allocates its output version ID at admission. It does not change the published version.

Inspect a candidate

const version = await rebyte.models.versions.retrieve(model.id, run.version_id);
for await (const item of rebyte.models.versions.list(model.id)) {
  console.log(item.id, item.status, item.optimizer_steps, item.expires_at);
}

Version states are queued, training, verifying, ready, failed and cancelled. ready means the version passed a serving check. It is not a claim that its answers are better for your task. Inspect its source, settings, training progress, context limit, output limit and pricing.

A version has a finite lifetime. Check the returned expires_at before using it for inference or as the source of another run. Publishing does not extend that lifetime. An expired or unavailable version fails explicitly.

Evaluate on held-out work

Use the exact version ID for evaluation:

const evaluation = await rebyte.sessions.create({
  scenario_id: scenario.id,
  agent: {
    model: run.version_id,
    instructions: 'Extract order facts as JSON.',
  },
  environment: { type: 'none' },
  input: 'Order B-210 contains 3 pens at $2 each. Return order_id and total.',
});

Review completed Turns and save feedback. Turns, feedback and traces retain the actual model_version_id, so later publication cannot change which version produced a result.

Keep evaluation Turns outside future training selections. Compare the same held-out tasks across versions and inspect errors, not only one success rate. Rebyte does not choose a winner or publish automatically.

Publish and roll back

await rebyte.models.publish(model.id, { version_id: run.version_id });

// New Sessions can now use the stable family ID.
const session = await rebyte.sessions.create({
  agent: { model: model.id, instructions: 'Extract order facts as JSON.' },
  scenario_id: scenario.id,
  environment: { type: 'none' },
  input: 'Order C-310 contains 4 folders at $3 each. Return order_id and total.',
});

curl --fail-with-body "https://api.rebyte.ai/v1/models/$MODEL_ID/publish" \
  -H "Authorization: Bearer $REBYTE_API_KEY" \
  -H 'Content-Type: application/json' \
  -d "$(jq -n --arg version "$VERSION_ID" '{version_id:$version}')"

Publication updates the family's current_version_id. New Sessions resolve it once and remain pinned for the conversation. Existing Sessions keep their original version, even if you publish or roll back later.

To roll back, publish an earlier owned, ready, unexpired version of the same family. To continue training, select that version as a new run's source. These are separate actions: rollback changes serving; a new training run creates a new experiment.

HTTP reference

Reads require tasks:read; writes require tasks:write. These extension routes need no OpenAI-Beta header. Lists use limit, order and after.

Create training through training runs. Model families and versions currently have no delete endpoint.