> ## Documentation Index > Fetch the complete documentation index at: https://docs.trajectory.ai/llms.txt > Use this file to discover all available pages before exploring further. # Download a checkpoint > Private beta: download a checkpoint's LoRA adapter or merged model. > [!NOTE] > > Private beta. Your organization needs access; otherwise these calls return 403 `beta_not_enabled`. Contact Trajectory to request it. These helpers may change in a minor release. Requires `trajectory-sdk` 0.10.11 or later. Download a [checkpoint](https://docs.trajectory.ai/guides/training.md) to use its weights outside Trajectory. You can download its LoRA adapter, or for supported base models, the full model merged with its base. ## Download with the SDK ```python theme={null} from trajectory import Client from trajectory.lib.beta import download_checkpoint client = Client() checkpoints = client.training.checkpoints.list("YOUR_TRAINING_RUN_ID") checkpoint = max(checkpoints, key=lambda c: c.step_index) path = download_checkpoint(client, checkpoint.checkpoint_id, "models/") print(path) # models// ``` The first request for a checkpoint prepares its export, which can take several minutes for a large adapter. Later requests reuse it. The directory holds the adapter files and `manifest.json`. | Argument | Default | Behavior | | - | - | - | | `output_dir` | `.` | Parent directory for `/`. | | `overwrite` | `False` | Replace an existing target. Otherwise `FileExistsError` is raised before any request. | | `timeout` | `3600` | Seconds to wait for the export to be ready. `None` waits indefinitely. | | `poll_interval` | `5` | Seconds between status checks. | | `http_client` | New client | Your own `httpx2.Client` for the download. | A failed call leaves nothing behind and starts over the next time. Downloads do not resume. ## Download from the platform Open the training run, select a checkpoint, and choose **Download weights**. The button appears for deployable checkpoints when your organization has access. ## Use the download link yourself `wait_for_checkpoint_export` returns the ready export with a signed link, without downloading anything: ```python theme={null} from trajectory.lib.beta import wait_for_checkpoint_export export = wait_for_checkpoint_export(client, "YOUR_CHECKPOINT_ID") print(export.download.url, export.download.size_bytes, export.download.url_expires_at) ``` The link lasts one hour. `client.beta.checkpoint_exports.retrieve(export.export_id)` returns a new one. `client.beta.checkpoint_exports` also has `create` and `list`, and sends the beta opt-in header for you. ## Handle errors | Error | Meaning | What to do | | - | - | - | | `trajectory.lib.beta.CheckpointExportError` | The export failed or has no download link. The error's `export` attribute holds it. | Call again to retry. | | `TimeoutError` | The wait ended first. The server keeps preparing the export. | Call again to pick it up. | | `zipfile.BadZipFile` | The downloaded ZIP is corrupt. | Call again to download it again. | | 403 `beta_not_enabled` | Your organization doesn't have access. | Contact Trajectory. | | 409 | This checkpoint can't be downloaded. | Only Tinker checkpoints are supported. | ## Download a merged model A merged model is the base model with the adapter merged in, stored as a Hugging Face folder of files. Merged exports work like a batch job: queue the export, poll it, and pull it down once it's ready. Merging can take hours, and a model can be hundreds of gigabytes, so nothing here needs a process that stays open. Your organization needs merged exports enabled separately. ```python theme={null} from trajectory import Client from trajectory.lib.beta import download_merged_export client = Client() # Queue it; this returns right away, and repeat calls reuse the same export export = client.beta.checkpoint_exports.create( "YOUR_CHECKPOINT_ID", body={"export_type": "merged_model", "precision": "bf16"} ) # Poll later, even from another process or machine export = client.beta.checkpoint_exports.retrieve(export.export_id) # Pull it down once it's ready if export.state == "ready": path = download_merged_export(client, export.export_id, "models/") print(path) # models/-merged-bf16/ ``` There's no webhook, and a queued export can't be cancelled. `download_merged_export` downloads files in parallel and checks each against its MD5. If a call fails, for example because a link expired during a long download, call it again: files it finished stay in a hidden `..partial` folder next to the target, and the next call downloads only the rest with fresh links. | Argument | Default | Behavior | | - | - | - | | `output_dir` | `.` | Parent directory for the model folder. | | `overwrite` | `False` | Replace an existing target. Otherwise `FileExistsError` is raised before any file request. | | `max_workers` | `8` | Files downloaded at once. | ### Supported models `client.beta.checkpoint_exports.list(checkpoint_id).available_merge_precisions` lists the formats a checkpoint can use, with each one's size and known limits. | Base model | Precision | Merged size | | - | - | - | | `Qwen/Qwen3.5-4B` | `bf16` | 9.3 GB | Exports expire after 28 days. A retrieved export then reports `failed`, and queueing it again rebuilds it. ### Merged export errors | Error | Meaning | What to do | | - | - | - | | `CheckpointExportError` | The export isn't a ready merged export. The error's `export` attribute holds it. | Check its state. | | `ValueError` | A file failed its MD5 check and wasn't kept. | Call again to fetch it. | | `requests.HTTPError` or `requests.ConnectionError` | A file failed to download, or its link expired. | Call again; finished files are kept. | | 409 | This base model or precision isn't available, or your organization doesn't have merged exports. | Check the supported models. | Evaluate a merged model on your own serving stack before relying on it, because merging can reduce accuracy, especially at low precision. ## Build a full model yourself For other base models, merging the adapter happens on your side. Tinker's [merge guide](https://tinker-docs.thinkingmachines.ai/cookbook/deployment/export-hf/) covers how, and its supported models are listed in the [Cookbook](https://github.com/thinking-machines-lab/tinker-cookbook/blob/main/tinker_cookbook/weights/README.md). Evaluate the result on your own serving stack, because merging can reduce accuracy, especially at low precision. This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.