Distillation Policy

NanoGPT exposes a distillationPolicy field that indicates whether a model or provider route is allowed for output-based model training or distillation under NanoGPT’s recorded model-license and provider-terms rules. This metadata is currently API-only. The text models browser UI does not show a distillation filter or badge. Distillation policy metadata is based on NanoGPT’s current interpretation of published model licenses and provider terms. Users are responsible for ensuring their use complies with all applicable terms and laws.

Policy Object

Text model records and provider rows may include:

{
  "distillationPolicy": {
    "status": "allowed",
    "label": "License permits distillation",
    "basis": "permissive-open-weights",
    "sourceUrl": "https://example.com/license-or-terms",
    "note": "Short explanation of the policy signal."
  }
}

All fields are additive and optional for backwards compatibility. New API responses include distillationPolicy where NanoGPT has recorded policy metadata.

Field	Type	Description
`status`	string	One of `allowed`, `disallowed`, or `unknown`.
`label`	string	Human-readable policy summary.
`basis`	string	Rule basis used to derive the status.
`sourceUrl`	string	License or provider terms URL used as the policy source, when recorded.
`note`	string	Short explanation and caveats for the policy signal.

Statuses

Status	Meaning
`allowed`	Distillation is allowed under the current recorded rule set.
`disallowed`	Distillation is restricted by model or provider terms.
`unknown`	NanoGPT does not have a clear enough signal.

Bases

Basis	Meaning
`deepseek-api-terms`	DeepSeek explicitly allows use of inputs and outputs for training other models, including distillation.
`permissive-open-weights`	The underlying model or license permits distillation or derivative model use.
`provider-output-use-terms`	The provider terms explicitly permit output use.
`provider-output-use-restriction`	The provider terms restrict output-based training, distillation, model development, or competing model/service creation.
`closed-provider-restriction`	The model is from a closed provider family that commonly restricts using outputs for competing model training.
`provider-terms-unknown`	No clear provider or model policy is recorded.

Model Policy vs Provider Policy

There are two policy layers:

Layer	Question it answers
Model/license policy	Does the underlying model or model family allow distillation?
Provider/API policy	Does the specific hosted provider route allow users to use API outputs for distillation?

A provider can restrict output use even when the underlying model license is permissive. Provider-specific restrictions override model-level allowances. Provider rows are evaluated with these rules:

Explicit provider restrictions return disallowed.
Explicit provider permission returns allowed.
If provider terms do not mention output-based training or distillation, and the underlying model/license allows it, NanoGPT treats the provider route as allowed.
If the underlying model policy is unknown or disallowed, provider silence does not make it allowed.

Example provider row:

{
  "provider": "fireworks",
  "distillationPolicy": {
    "status": "allowed",
    "label": "Provider terms permit output use",
    "basis": "provider-output-use-terms",
    "sourceUrl": "https://fireworks.ai/terms-of-service",
    "note": "Fireworks permits use of outputs for any lawful purpose, subject to model-provider license and usage restrictions. Model-level signal: license permits distillation."
  }
}

Filtering Text Models

The explore text model endpoints support distillation=allowed.

GET /api/explore/text-models?distillation=allowed

This returns only text models whose model-level distillation policy is allowed. Search supports the same parameter for text search:

GET /api/explore/search?type=text&q=qwen&distillation=allowed

The filter applies only to text models.

Response Metadata

When the filter is used, responses include:

{
  "meta": {
    "distillation": "allowed"
  }
}

Without the filter, text model responses include:

{
  "meta": {
    "distillation": "all"
  }
}

Provider-Level Behavior

Provider route policy can differ from model-level policy:

DeepSeek models served by DeepSeek are allowed.
Fireworks, Together, Nebius, and ArliAI routes on a model whose license permits distillation are allowed.
Providers with explicit competing-model or output-training restrictions are disallowed.
Providers with no recorded restriction on an allowed open-weight model are allowed.
Providers with no recorded restriction on an unknown model or license are unknown.

For a representative open Qwen model across user-selectable providers, the current provider-level split is 35 allowed, 21 disallowed, and 0 unknown.

Caveats

This metadata is informational and is not legal advice.
Provider terms and model licenses can change.
Model licenses can impose attribution, naming, acceptable-use, or derivative-model restrictions even when distillation is allowed.
Provider-specific restrictions override model-level allowances.
Unknown model/license status remains unknown even if a provider is silent.
Inspect sourceUrl and note before relying on a model or route for distillation.

Documentation Index

​Distillation Policy

​Policy Object

​Statuses

​Bases

​Model Policy vs Provider Policy

​Filtering Text Models

​Response Metadata

​Provider-Level Behavior

​Caveats