Skip to main content
Register, upload, and inspect the models and adapters you deploy with dedicated model inference. Commands run within a Together project. Bringing your own weights is a two-step flow:
  1. create registers a model record (metadata only).
  2. upload or remote-uploads create adds the weight files.
To deploy a registered model, run tg endpoints deploy or its alias, tg models deploy.
These commands require Together CLI 3.0 or later. In CLI 2.x they were tg beta models, and tg models listed the serverless catalog and uploaded your own models for the retired v1 dedicated endpoints.

Create

Register a new model record. This creates metadata only. It does not upload any files. Upload the weights afterward with upload or remote-uploads create.
Shell
You can also pass a base model ID (ml_...). Values that do not start with ml_ are treated as a deploy model name: the CLI searches public supported models and requires an exact match on a deployment profile’s modelName. If the name is missing or matches more than one profile, the CLI prints the candidate model names and ids and asks you to re-run with an exact name or ml_... id. Alias: tg models -c.

Parameters

Upload

Upload weight files from your local machine to an existing model or adapter record.
Shell

Parameters

List

List the models in the current project.
Shell
Alias: tg models ls.

Parameters

Retrieve

Get a single model by ID.
Shell
The CLI looks for the model in the current project first. If that lookup returns a 404, the CLI retries against the internal-visibility models in your organization, which you can list with tg models org. Any other error, such as a 403, is returned as-is. For a model found in the organization, the CLI reads the file list from the project that owns the model. If the organization lookup also returns a 404, the CLI returns the not-found error from the project lookup. Alias: tg models get.

Update

Update a model record. Only the fields you pass are changed.
Shell

Parameters

Delete

Delete a model record. This removes only the metadata, not the uploaded files.
Shell
Alias: tg models -d.

Configs

To deploy a model with a specific config, pass its config revision ID (cr_...) to tg endpoints deploy --config.

Configs list

List the deployable configs for a model. Pass a bare model ID (ml_...), a resource path (projects/{project_id}/models/{model_id}), or a model name.
Shell
For a model name with a /, the CLI searches your project’s models when its prefix matches your project slug, and public supported models otherwise. If a public name matches more than one model, the CLI prints an error that points you to tg models public --search. Alias: tg models configs ls.

Parameters

Configs get

Get a config revision by ID. You can leave out get and pass the config revision ID (cr_...) straight to tg models configs.
Shell
The output shows the config’s reference model, its draft model if it has one, its selectors, and the models it’s certified for.

Parameters

List files

List the files in a model or adapter.
Shell

Parameters

List revisions

List the revisions of a model, including each revision’s validation status.
Shell
Alias: tg models list-revisions.

Parameters

Download

Download the weight files of a model or adapter to a local directory.
Shell

Parameters

List public models

List the publicly visible models across all projects. Use this to find a base model ID or deployment profile modelName for create and tg endpoints deploy.
Shell
The human-readable table prints one row per deployment profile, with columns for Model (the profile’s modelName, such as Qwen/Qwen3.5-9B-FP8), GPUs, and Parallelism. Pass that Model value as the MODEL argument to tg endpoints deploy when you want that specific profile. Add --json to see each architecture’s full record, including its id (which subsequent operations require) and its deploymentProfiles array. A model can expose more than one profile, each pairing a certified config with a hardware and parallelism choice, so read the profile’s modelName, config, gpuType, and gpuCount before deploying.

Parameters

List org models

List the internal-visibility models in your organization.
Shell

Parameters

Remote uploads

Import model weights server-side, without downloading them locally first. The source can be a Hugging Face repo or a presigned https:// URL on Amazon S3, Google Cloud Storage (GCS), or Cloudflare R2. Create a remote upload job against an existing model record, then poll it until it succeeds.

Remote uploads create

Start a remote upload job.
Shell
For gated or private Hugging Face repos, also pass your Hugging Face token with --token. For S3, GCS, or R2, pass a presigned archive URL as --from, with no token. Other sources are rejected. To import from an s3:// path, presign the object and pass the presigned URL instead. Alias: tg models remote-uploads -c.

Parameters

Remote uploads retrieve

Get a remote upload job by ID. Poll the job until its status is REMOTE_UPLOAD_STATUS_SUCCEEDED.
Shell
Alias: tg models remote-uploads get.

Parameters

Remote uploads list

List the remote upload jobs in the current project.
Shell
Alias: tg models remote-uploads ls.

Parameters

Global options

Every command also accepts the global parameters, including --json for machine-readable output and --project to override the target project.

Resource IDs