Import from a provider

Import from a provider

1. 1

   Connect

   Enter the provider's API base URL and an optional API key, then Fetch models. obleth reads GET {base}/models from the gateway, so the key never reaches your browser. A 404 usually means the base should gain or drop its /v1.

2. 2

   Select

   Every model the provider serves is listed. New ones are selected; ones you already have are hidden behind Show N already imported and can't be picked. Set the defaults for every model (type, context window, input and output cost per token), then change each row's name or type where you need to.

3. 3

   Review

   Review import shows each model as it would be created, with any problem for every model at once, such as an alias already taken. Confirm to create them.

Imported models point at the provider's base URL and ID and carry the key you gave. Like a model added by hand, each is created switched off and turns on when its first health check passes; the result says which came on and why the others stayed off. Run the import again later and only new models are offered.

Names and formats

Each provider ID becomes a suggested name: lowercase, with anything outside a-z 0-9 . - turned into dashes. A trailing format token (fp8, bf16, fp16, mxfp4, nvfp4, int8, int4, awq, gptq, gguf) moves into the model's quantization, and the full name becomes an alias, so callers of the provider's spelling keep working:

The model's upstream model is the provider's ID, unchanged. Two selected models that end up with the same name are refused until you rename one.

A provider model counts as already imported when one of your models has its name (with or without the format suffix), or has the same base URL and upstream ID. That second check keeps a model you renamed after importing it from being offered again.

Import a file

Import or export on Models writes models to an obleth-models manifest and reads one back, on the same gateway or another. Use it to keep models in a repository or copy them between gateways. Import a file shows what would be created and changed before anything is written. A manifest is not a backup: for a whole-gateway copy, see Backup and restore.

curl -s http://localhost:9180/api/v1/models/export \
  -H "Authorization: Bearer $TOKEN" > obleth-models.json

# Preview without writing, then apply
curl -s -X POST "http://localhost:9180/api/v1/models/import?dry_run=true" \
  -H "Authorization: Bearer $TOKEN" -H "Content-Type: application/json" \
  --data @obleth-models.json
curl -s -X POST http://localhost:9180/api/v1/models/import \
  -H "Authorization: Bearer $TOKEN" -H "Content-Type: application/json" \
  --data @obleth-models.json

format: obleth-models
version: 1
models:
  - model_name: glm-5-3
    aliases: [glm-5-3-mxfp4]
    upstream_model: zai-org/GLM-5.3-MXFP4
    api_base: https://inference.example.com/v1
    quantization: mxfp4
    context_window: 131072
    tags: [coding, "reasoning:3"]
    endpoints:
      - name: primary
        api_base: https://inference.example.com/v1
        priority: 100
        weight: 100
  # A sparse entry changes only what it names.
  - model_name: qwen3-embedding-8b
    tags: [general]

Models are matched by model_name: an existing name is updated and a new one created. Every other field is optional, and a field left out is left unchanged. Import never deletes a model or an endpoint; endpoints are matched by name. Unlike a provider import, a model the file adds takes the on or off state the file gives it (on when the file doesn't say).