Use your fine-tuned model

Deploy a ready version in the console, then send its name as model on /v1/decide: decisionnode-1.0-flash@acme/fraud for the version deployed most recently, decisionnode-1.0-flash@acme/fraud:3 for one version. The response's model always names the version that answered.

Deploy#

Open the version in Fine-tuning and choose Deploy. It shows deploying and then deployed, from which moment its names answer requests; if it shows deploy failed, deploy it again. Up to 3 versions can be deployed at once in a workspace. Undeploy stops a version: from that moment its names are unknown to the API.

The name#

Grammar

<base>@<workspace>/<name>[:<version>]

The parts of a fine-tuned model's name
Part	Rule	Example
`<base>`	The pinned id of the model it was trained from	`decisionnode-1.0-flash`
`<workspace>`	Your workspace's slug, as the console shows it	`acme`
`<name>`	The model's name, 3 to 40 of `a-z`, `0-9` and `-`	`fraud`
`:<version>`	Optional: one version. Without it, the version deployed most recently	`:3`

Send the plain name (decisionnode-1.0-flash@acme/fraud) to follow your deployments: when you deploy version 4, the same name answers from it.
Send a pinned name (decisionnode-1.0-flash@acme/fraud:3) where answers must not change: it answers from that version as long as it is deployed.
Log the response's model. It is always the pinned name, also in the x-decisionnode-model header, so you know which version decided.

Call it#

curl https://api.decisionnode.com/v1/decide \
  -H "Authorization: Bearer $DECISIONNODE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "decisionnode-1.0-flash@acme/fraud",
    "state": {
      "claim": {
        "merchant": "Taxi Stockholm",
        "total": "512.40",
        "currency": "SEK",
        "receipt_total": "312.40",
        "prior_claims_90d": 1
      }
    },
    "questions": {
      "fraud": {
        "type": "truth",
        "instructions": "Is this claim fraudulent?"
      },
      "route": {
        "type": "choice",
        "instructions": "What should happen to this claim?",
        "criteria": {
          "pay": "pay the claim in full",
          "partial": "pay only the part the receipt documents",
          "decline": "refuse the claim"
        }
      },
      "risk": {
        "type": "score",
        "instructions": "How risky is this claim?",
        "criteria": ["none", "low", "medium", "high"]
      }
    }
  }'

Response

{
  "model": "decisionnode-1.0-flash@acme/fraud:3",
  "answers": {
    "fraud": { "type": "truth", "truth": 0.08 },
    "route": {
      "type": "choice",
      "choice": "partial",
      "confidence": 0.86,
      "probabilities": { "decline": 0.04, "partial": 0.91, "pay": 0.05 }
    },
    "risk": {
      "type": "score",
      "score": 1.83,
      "confidence": 0.71,
      "legend": { "0": "none", "1": "low", "2": "medium", "3": "high" },
      "probabilities": { "0": 0.02, "1": 0.2, "2": 0.71, "3": 0.07 }
    }
  },
  "usage": { "input_tokens": 118, "output_tokens": 0 }
}

The request is any request the base takes: the same questions, images, videos and boosters. Ask the questions the model was trained on, in the same words, for the gain the gate measured; other questions are answered too, without the gate's measure behind them.

List your models#

GET /v1/models lists your workspace's fine-tuned models beside ours, for your workspace's keys only: the plain name of every model with a deployed version, and every deployed or ready version under its pinned name. Only deployed entries answer.

JSON

{
  "models": [
    {
      "name": "decisionnode-1.0-flash@acme/fraud",
      "description": "Claims fraud, trained on 2,400 decisions",
      "release_date": "2026-10-09",
      "version": "decisionnode-1.0-flash@acme/fraud:3",
      "fine_tuned": true,
      "base": "decisionnode-1.0-flash",
      "adapter_version": 3,
      "status": "deployed"
    },
    {
      "name": "decisionnode-1.0-flash@acme/fraud:3",
      "description": "Claims fraud, trained on 2,400 decisions",
      "release_date": "2026-10-09",
      "version": "decisionnode-1.0-flash@acme/fraud:3",
      "fine_tuned": true,
      "base": "decisionnode-1.0-flash",
      "adapter_version": 3,
      "status": "deployed"
    },
    {
      "name": "decisionnode-1.0-flash@acme/fraud:4",
      "description": "Claims fraud, trained on 2,400 decisions",
      "release_date": "2026-10-09",
      "version": "decisionnode-1.0-flash@acme/fraud:4",
      "fine_tuned": true,
      "base": "decisionnode-1.0-flash",
      "adapter_version": 4,
      "status": "ready"
    }
  ]
}

The fields of a fine-tuned entry

fine_tunedboolean: true. Our own models' entries do not carry it.
basestring: The base's pinned id.
adapter_versioninteger: The version number; on a plain name, the version it answers from.
statusstring: deployed, or ready: trained and passed the gate, but not deployed, so its name is refused.
versionstring: As on every entry: the pinned name this entry answers from.

Errors#

What a fine-tuned name can get back
Status	Type	When
`400`	`api_usage_error`	`Unknown model: <name>`: a name that does not exist, a version that is not deployed (a `ready` one included), or another workspace's model. The answer is the same in every case and never says whether a name exists elsewhere
`529`	`overloaded_error`	The model has no ready capacity, typically on its first request after a quiet spell: wait `Retry-After`, then retry. It never falls back to another model
Any other	As on the base	A fine-tuned model answers every other error exactly as its base does; see Errors

When it has no ready capacity#

A fine-tuned model that has not been called for a while can take a few minutes to become ready again. Until it is, its requests answer 529 overloaded_error with Retry-After: wait that long and retry, as you would any 529 (the SDKs do it for you). A fine-tuned model never falls back to its base or to DecisionNode-⁠1.0 Flash: you asked for your own model, so nothing else answers in its place, and the fallback field has no effect on it.

Expect it after a quiet spell, not under steady traffic: a model that keeps receiving requests stays ready.
Retry from a queue, not inline, where a decision can wait a few minutes; for one that cannot, decide with the base model meanwhile and log which model answered.

Price#

A request to a fine-tuned model costs its base's price per input token plus $0.02 per million input tokens: $0.062 on DecisionNode-⁠1.0 and $0.055 on DecisionNode-⁠1.0 Flash at the proposed base prices. Output stays free. Usage shows each fine-tuned model under its pinned name. See Limits and pricing.