Manage models in the current Link window's Models tab, during setup, or from Control. A model being downloaded, ready to serve and loaded in memory are different states.
With Link 2.4.0’s routing flow, Control can select a local model as the final backup. Link downloads and prepares it automatically. Backup · Loads when your first model can’t answer is different from ordinary serving; you do not need to select Start serving to prepare that backup.
When no local model serves normally, Models can show Serving through the configured cloud model. Requests sent through that route leave this device. Inspect Routing for the applied order and use Manage in Control to edit it.
Models distinguishes the active cloud destination from a prepared local backup. Model details provides the Model ID; Serving settings provides context and memory controls.
Find the section
Check the serving destination
The outlined area is shown close up on the next slide.
Close-up
Check the serving destination
Models shows the cloud destination in use and a local model marked Backup. A prepared backup loads when needed; Start serving is a separate local-serving control.
Slide 04
Copy the Model ID
Expand Model details and use Copy beside Model ID for a manual client. The downloaded file and its size appear below. Remove download removes this device copy.
Slide 05
Review context and memory
Auto selected 64K on this device, with a 2.8 GB estimate. Keep loaded can reduce the wait for a first request while using memory. Values depend on the model and hardware.
Slide 06
Apply deliberate memory changes
Advanced memory settings contains Fit more context, using 8-bit context memory with a possible quality trade-off. Select Apply after a change. No change is pending in this capture, so Apply is disabled.
Slide 1 of 6: Manage local models and backups.
Aventail Link 2.4.1 on macOS · captured 14 November 2026. Account/device identity obscured. Model names and memory values are examples. Use the arrows or progress marks to move between slides.Read the current Link model controls
Open Models → Add model and choose a model available to your account and device. Wait for its download to finish.
Select Start serving beside the downloaded model. Wait for Ready to serve. Unless kept loaded, it loads when the first request arrives.
Expand Serving settings to inspect Context length and Estimated memory. Auto chooses a context that fits the device; longer context requires more memory. The captured 64K value is an example, not a requirement for every device.
Under Advanced memory settings, Fit more context uses an 8-bit context cache where supported. Inspect the memory estimate and quality requirements before applying a change.
Use Keep loaded when you want to reduce the wait for the first request, allowing for memory use and other models. Select Apply after changing serving settings. Changes can reload the model, so finish important work first.
Select Stop serving after active work finishes if the model should no longer answer requests. Stopping serving leaves the downloaded model on the device.
Use the download's Cancel download or Retry download controls when appropriate. Preserve a reported failure reason and check disk space, connectivity and account download capacity before trying again. A model that partly uses the CPU may respond more slowly; read the reason shown rather than assuming the accelerator is fully used.
Finish active requests before opening the intended local model’s Model details and choosing Remove download. Link 2.3.2 waits for its engine to stop and its files to be removed, and reports a failure if that cannot complete. Shared model files still needed by another model are retained. Removing this device copy is different from deleting the account's library model.
In Link, open Models, select the model, expand Model details and use Copy beside Model ID (called Model identifier in 2.3.2). Copy the identifier rather than typing the model’s display name: the two can differ. You can also copy Link’s address under Connected apps → Connect another app yourself, with Copy address.
aventail models serve MODEL --wait aventail models stop MODEL
Replace MODEL with the installed model's identifier or serving alias. For language models, --wait waits for the model to be loaded and answering, rather than only enabling its serving configuration. A timeout or error needs investigation before retrying. Use aventail models --help for the current deploy, set and remove options.
The earlier 2.0.0 Preferences controls are retained below with their version context.
Earlier Aventail Link 2.0.0 model-management reference
AVENTAIL / LINK
Manage device models
1 / 5
Full view
Model deployment in Control
Select Deploy on a model to choose the device that will download it.
Find the section
Target device and Deploy action
The outlined area is shown close up on the next slide.
Close-up
Deploy from Control
Choose a compatible model and the online device. Wait for deployment to finish before serving.
Slide 04
Download another model
Choose a model under Available to download and select Download. Wait for installation to finish before enabling serving.
Slide 05
Manage installed models
Use the model checkbox to enable or disable serving. Read the model details and status; Test and Remove are on the same row.
Slide 1 of 5: Model deployment in Control.
Control deployment and Aventail Link 2.0.0 Preferences. The Preferences image is a redacted copy supplied on 2 October 2026. Use the arrows or progress marks to move between slides.
Open Preferences… from the menu-bar icon. Under Available to download, choose a model and select Download. Wait for installation to finish, then use its checkbox under Models to enable serving. Check the status beside the model; Test is also available on the row.
Model availability depends on account access and the installed runtime. If a model is absent, check the catalogue in Control rather than manually copying files into Link's folders.
Finish active requests and clear the model's serving checkbox. In Preferences → Models, identify the model and use Remove. Read any confirmation and check the device's model list afterward.
If a download is stalled, record its status and error, check connectivity and disk space, then use the available cancel/retry controls. Do not delete runtime files manually as a first recovery step.