Prepare and verify
Prepare turns local source weights into an MLX model. Every conversion is previewed before it runs, and supported task types are verified before they are marked ready.
Preview a conversion
- Choose a source: from Library (Prepare to run), LM Studio, the Add from Hugging Face window, or Choose in Library.
- Pick a quantization width. Each tile shows the output size and fit on this Mac above a Source-to-Ready pipeline track; a Hugging Face source uses the task the intake window labelled. The default width comes from Settings → Conversion defaults.
- Preview conversion asks the engine for a plan: the exact command, output path and a preview hash. Choosing a different width re-targets the destination.
- Choose whether to move the original source to the Trash after verification. The choice you see is the one recorded. A source outside the model folders in Settings cannot be moved, so the choice is off and says the original stays.
- Confirm conversion runs exactly the reviewed plan.
Prepare follows the job while it runs: a progress ring tracks bytes written against the intake estimate (indeterminate while the converter loads), and the job log sits in a collapsible section. Only one conversion runs at a time.
The quality gate
After a fresh scan confirms the output, the conversion quality gate runs the canary for the model's task type:
| Task type | Canary |
|---|---|
| Text LLM, vision-language | Served on a temporary loopback port with reasoning switched off where the chat template allows; answers canary prompts, scored on the answer after any <think> block |
| Speech-to-text | Transcribes a sentence synthesized with say; passes at 25% word error or less |
| Classification | Answers a billing ticket that must route to billing and read as a refund request |
| Image generation | Renders a fixed prompt at 512×512 and must write a non-blank image |
A pass marks the workflow verified only after its report is saved; a failure marks it verification failed. If the report cannot be saved, the output remains complete but unverified, the storage error is shown and no new verdict is published. Resolve the storage problem before retrying verification. Keep anyway also requires its override to be saved before changing the workflow. Task types without a canary complete without one. Turn the gate off in Settings → Advisors → Verify conversions.
A notification announces each terminal state: completed, verified, verification failed or failed. Restoring state at launch never notifies.
Activity
Activity puts active comparisons and in-flight conversions before older records. Study and training work that is queued, running, failed or in an unknown state stays prominent; successful and cancelled jobs remain in expandable history. Failed conversions appear under Needs attention. Other conversion records stay grouped by creation day. Each conversion has a stage track, its state and one action.
Empty study and training sections are omitted. Their status failures stay visible with the latest available records, rather than appearing as empty history. Open Studies and Open Training lead to their full workflows.
Refresh reads conversion, serving, study, training and queue status without starting work. A successful conversion read does not clear a training failure. If serving status is unavailable, existing server rows are labeled as last known and Activity does not claim that no server is running.
- Confirmed conversions are written to a durable queue before launch and drain first in, first out.
- Actual running state comes from engine receipts, never from the queue alone.
- Run next reconciles receipts and starts one pending item when compute is free.
- Retry reuses the reviewed plan. Moving a queued item with ↑ and ↓ does not approve a new plan.
- Failed launches keep their error and remediation.
The engine owns the shared conversion queue. It is collapsed when empty; queued records or a queue error keep it visible. Import existing queue adopts a convert-queue.json from an earlier Workbench installation; stop the former server first. Import preserves the source file and an exact backup, and refuses invalid state.
Clean up the source
A verified conversion can reclaim the files it was made from. Prepare's Preview source cleanup and Reclaim's Original sources after verification list eligible originals: receipt-owned GGUF files, or Hugging Face cache repositories with a single revision. Changed outputs, active or preferred models, other revisions and blobs shared with other models are protected. Move originals to Trash moves only what the preview showed. See Reclaim for restore.
Run from Prepare
A verified chat or vision-language model offers Run model. Choosing it opens Run with that model selected and leaves the conversion record unchanged. Non-servable task types do not offer Run.