Library and Overview

Overview

Overview is the landing screen. It leads with a Model budget instrument for this Mac, offers one Next step action and shows the Flight path of the model chosen in Library: how far it has come from source to ready. Watch alerts appear here; see Health and alerts.

The Next step follows the model selected in Library: ready chat models open Run for a serve preview, sources needing conversion open Prepare, and other task types open their own Library inspector. Missing runtimes lead to repair guidance. An unavailable selection asks you to choose another model. Active conversions, verification and workflow failures still take priority. While a scan is running or has failed, resolve the scan before using selected-model guidance. These actions navigate or stage a preview; they never start conversion, installation or serving.

Library

Library lists the GGUF and MLX models under your configured roots. The output directory is scanned as an MLX root unless another root already reaches it. Families with multiple variants expand to show their models; single variants appear directly in the table.

The Source menu offers All local models and each saved model host, including Ollama, LM Studio and ComfyUI. Host choices use their saved identity and show their provider; connections with matching names and providers include their address to distinguish them. Adding, editing or removing a saved connection updates the menu. Hosted models opens host selection and connection setup. The separate Local files group contains LM Studio files.

Switching sources preserves the main table's filters and selected model without refreshing a host, changing configuration or starting conversion or serving. Selecting a host shows that exact connection. If it is removed, Library returns to hosted selection rather than displaying another host.

Show in Library on a source model opens All local models, clears search and readiness/quantization filters, and selects the model in the inspector. This explicit reveal leaves column and grouping preferences intact.

Columns collapse as the window narrows. Filters narrow the list by quantization and readiness. Refresh rescans; a failed scan offers Retry scan. Model-folder metadata is read in the background. The previous inventory stays visible until the complete replacement is ready; a cancelled refresh cannot publish its result. Repeated refresh requests do not start overlapping scans. The last successful inventory stays visible during retries. A retry clears the previous scan error while it runs. If scanning times out repeatedly, check the model folders in Settings for unavailable drives or unnecessarily broad folders.

When a scan finds no models, Add from Hugging Face… opens intake for review, Open Settings lets you choose model folders, and Scan again refreshes the inventory. Opening intake does not download a model. If search or filters hide existing models, Clear search and Clear filters reveal them without adding another model.

Model details

In wide windows, the first selected model opens the inspector beside the table. In compact windows, selecting a model opens its details in a sheet. Use Inspector in the toolbar or ⌥⌘I to hide or show details. Actions and the current memory fit come first; expand Model details for paths, runtime and other technical facts, including a speculative-decoding drafter's target. The inspector also holds:

• Repair runtime…, for models whose declared backend is missing: review and explicitly confirm its installation in Health, then Return to model to refresh readiness and reveal the same model. No model download or execution starts from this navigation.

• Model lineage: a provenance timeline of source, converted, verified, benchmarked, served, wired and quarantined events, built from the records the app already keeps. Events whose model signature no longer matches are dimmed. Export it as Markdown or JSON.
• Measured performance: throughput and first-token latency aggregated from past comparisons.
• Generate image, for ready image-generation models: enter a prompt and use the displayed defaults, or adjust dimensions, steps and seed in Advanced options. Completed output is retained with its task and can be inspected or exported as a verified PNG.
• Understand image, for ready vision-language models: choose a local image and ask a question. The engine retains a verified input copy with the task; the result shows the full answer and input image.
• Classify, for ready classification models: a state description and questions with choice, score or probability answers.
• Generate music, for ready music models: caption, optional lyrics, duration, steps and seed, with playback of completed audio.
• Transcribe audio and Speak text, for ready speech models: transcription from a selected local audio file, or text-to-speech with optional voice, speed and language settings.

Actions: Run model opens the existing server preview flow; Compare selects the model for comparison. Models needing conversion offer Prepare to run. Verify now runs the canary suite for a ready model; when a run cannot reach a verdict, the Verification section says why. Reveal in Finder and Copy Path are in the model's action menu.

Select for Compare is available for ready chat, speech, music, image and video models. It opens the matching comparison mode and selects the model without starting a run. Run remains a chat-serving action. Runtime restrictions take precedence over task labels: Strata supports Chat only.

Local inference tasks

Choose a ready model and fill its task form in the inspector. Start submits one task to the engine and shows its ID and state. Editing inputs does not run the model. If the engine is busy or its backend is unavailable, the form keeps your inputs and shows the error. If the selected engine does not support the task, Open Health… takes you to the setup assistant action. In the assistant, Use bundled engine… offers an explicit preparation and confirmation step. Your task inputs stay in the form; try Start again after selecting a compatible engine. Passing a model's verification check does not establish that an older engine supports durable tasks for that model.

For Classify, describe the State and add at least one question with instructions. Choice questions use two or more ordered labels; score questions use two or more ordered level descriptions. A probability question (noul) can include descriptions of its false and true outcomes. Results show the model's returned answers, probabilities and score legend when available. Missing fields remain unavailable. These are model predictions, without ground-truth checks or an automatic quality verdict.

For Generate music, enter a caption and optional lyrics. Blank lyrics mean instrumental music. Defaults are 15 seconds, 30 steps and seed 42. The engine accepts captions up to 2,000 characters, lyrics up to 10,000, duration above zero through 360 seconds, 1–30 steps and seeds from 0 through 4,294,967,295. Completed tasks offer WAV playback and recorded generation settings. A valid audio file does not establish musical quality. Playback stops when you leave the inspector or select another model or task.

For Transcribe audio, choose a local WAV, FLAC or MP3 file and optionally enter a language code. The engine records a verified copy with the task before accepting it; later reuse reads that retained copy. Completed results show the full transcript and the input's recorded digest.

For Speak text, enter up to 2,000 characters and optionally set a voice, speed (0.5–2) or language code. Leave those controls blank to let the engine apply its defaults. A completed result exposes a verified WAV for playback and export. Workbench does not assign an automatic quality score to speech output.

For Generate image, enter a prompt of up to 2,000 characters. New requests default to 1024 × 1024 pixels, 40 steps and seed 42, shown above Advanced options. That disclosure lets you set width and height (multiples of 16 from 256 to 2048), steps (1–100) or seed. Clearing a setting leaves it to the engine planner. The completed task shows the verified PNG and supports PNG export. For Understand image, choose a local image and ask a question of up to 4,000 characters. Advanced options sets max tokens (1–4096), temperature (0–2) and max pixels (up to 100,000,000); blank values leave these to the engine planner. The engine records a retained image copy before accepting the task. The completed view shows that verified image and the full returned answer. Both task kinds appear in Recent tasks and can reuse their recorded request when the matching Ready model is available.

For Generate video, enter a prompt of up to 2,000 characters and optionally set width and height (64–1920), frame count (1–241), whole frames per second (1–60), steps (1–100) or seed. Blank settings are left to the engine planner, which also checks the model's size alignment and frame grouping. The completed task offers verified MP4 playback and export. For Understand video, choose a local MP4, MOV or M4V file and ask about its contents. Advanced options set frames per second (0.1–8), temperature, maximum tokens and maximum pixels, with the image limits above; blank values leave these to the engine planner. The engine retains a verified copy of the input with the task, and completed results show that video and the full returned answer. Video tasks appear in Recent tasks and can reuse their recorded request when the matching Ready model is available. Video playback stops when you leave the inspector or select another model or task.

Each form names an out-of-range value and keeps Run unavailable until it is fixed. Compare prompt sets apply the same limits.

Recent tasks in the inspector shows the selected model's tasks. The Library toolbar also opens Recent tasks independently of model selection, so recorded results remain accessible when their model is no longer in the Library. At narrow widths, find Recent tasks in the toolbar's overflow menu. The list contains at most the engine's 100 most recent records. Failed refreshes label the last known data and offer Retry.

The engine retains accepted inputs, task records and output files; agents using the same engine store can inspect the same task IDs. Closing Workbench does not cancel a task. Reopening refreshes its state without starting it again. Cancel requests that the engine stop the identified task; its row remains active until the engine reports a terminal state. If Start loses its response, Start outcome unknown offers Refresh without automatically resubmitting the request or guessing its task ID.

Reuse verifies the retained request and fills a new draft without changing the original task or starting another one. It requires the matching model to be ready and unchanged. Missing or changed request files prevent reuse; recorded results remain readable. Result and audio files are checked before display, playback or export. Missing or changed outputs show an error and do not trigger generation.

Result JSON expands the complete recorded payload. Export result saves verified result data and metrics while omitting request, source-path and artifact references; Export WAV saves a separately verified audio output. Private input content is not included by default. Workbench does not delete engine task artifacts, and Compare's ten-run media retention does not apply to these tasks. Use Compare for matched model runs and listening ratings.

Delete a model

Delete… moves a model to the Trash after a review that rechecks its identity, its folders and whether it is serving. Trash keeps the disk space until you empty it. For quarantine with restore, use Reclaim.

Task types

The engine labels each model with a task type. The type decides what the app offers.

Drafters such as llama.cpp dflash or eagle3 GGUFs borrow their target's embeddings and output head, so they are never served on their own.

LM Studio files

Choose LM Studio files in the Source menu's Local files group to list models in LM Studio's downloads folder from the last Library scan; it does not scan on its own. This filesystem shelf is separate from a saved LM Studio server connection. The folder comes from downloadsFolder in ~/.lmstudio/settings.json. The view states, per format, whether the last scan reaches that folder. Existing saved LM Studio navigation opens this shelf inside Library.

Two lanes are ranked by fit:

• MLX: models that run as they are; Run… opens Run.
• GGUF: files that need converting; Prepare… opens Prepare, or Run… when a converted output already exists. Split GGUF files are summed; vision projector files appear as tags.

The fit summary shows fit unknown until the MLX models have fit estimates. A partial set of estimates is labeled known to fit; a complete set uses fit now. Zero is shown only as the known count, never as a substitute for unavailable estimates.

The screen never writes configuration. When the scan does not reach the folder, add it to your roots in Settings.

Hosted models

Choose Library → Source → Hosted models, then Add host…. Presets cover Ollama, LM Studio, ComfyUI and llama.cpp; OpenAI-compatible accepts other local servers exposing a model-list API. Enter the server's HTTP or HTTPS loopback address, including its port. A trailing /v1 is accepted. Remote addresses are not supported. Optional API tokens are stored in macOS Keychain, separately from the saved connection.

The editor shows invalid names or addresses before Save. Test connection is optional: it reads the selected provider's model inventory using the entered replacement token, or the saved token when replacement is off. Authentication and missing-API errors offer guidance without showing server response bodies. A successful check shows the reported model count and check time; it does not test generation or establish that a model is ready to run. Editing any connection field or token clears the result. Cancel check stops waiting for a result. Testing never saves the host, changes Keychain, loads a model or publishes the inventory to Library. Save host remains the explicit persistence action; you can save a valid connection without testing it, then use Refresh later.

Refresh reads inventory without loading models, generating outputs, downloading weights or changing the host's configuration. Refresh is manual. A failed refresh retains the previous inventory with its capture time and an error. Editing an address or provider clears that connection's cached inventory. Removing a connection leaves the host and its models untouched.

Failed refreshes distinguish authentication, missing API and connection errors. Retry refresh rereads only the inventory. Edit connection… opens the saved provider, address and token controls; Cancel preserves the connection and error. When no prior inventory exists, the error says so. Retained inventory is dated evidence and does not establish the host's current state. A successful refresh replaces it and clears the error. Recovery actions never start generation.

The table shows model identifiers, available or loaded state, parameter size, quantization and file size where reported. Select a row and expand Reported model details for context, memory, capabilities and digest. Missing metadata is Unknown. Reported model memory is separate from whole-device memory; model context limits are separate from loaded context.

• Ollama lists installed tags separately from Reported running models. Expand that disclosure for each running model's digest, host-reported memory and loaded context, including weights absent from the installed tags or with a different digest. Warnings identify these differences. Running allocation size is not disk size. Maximum context and capabilities require an explicit model-details read.
• LM Studio uses its native inventory API, with compatibility fallbacks for older servers. Multiple loaded instances with different contexts have an unknown loaded context.
• ComfyUI lists model files by category, plus reported device and version information. File availability does not establish whether a model is loaded. Unsupported categories can produce a partial inventory with a warning.
• llama.cpp and compatible servers expose their advertised models. Additional metadata is shown only when reported; names never establish capabilities or identical weights.

For ComfyUI, expand Reported device memory in the host inventory or the Memory panel's Saved model hosts section. Refresh reads each device's reported free and total bytes with a separate capture time. Missing or invalid values are Unknown; a reported zero free-memory value remains zero, and an empty device list says No devices reported. Opening the disclosure uses cached readings without contacting the host.

MPS and CPU readings describe shared system memory and overlap system headroom. Other device types retain their host-reported readings. Devices are never added together; these values do not measure per-model allocation or memory released by cleanup. A reading describes its capture time, independently of inventory and queue activity.

For an installed Ollama tag, select its row, expand Reported model details, then choose Read model details. This reads the selected tag's reported capabilities and its architecture's maximum context, with a separate read time. Loaded context remains the current running allocation. Missing fields are Unknown; an explicitly empty capability list is None reported. The action requires a reported digest and checks that the installed digest matches before and after the metadata read. These are dated host observations, not an atomic runtime snapshot.

The read refreshes inventory and publishes the selected model's details only after the final identity check. Failed reads retain the previous dated inventory. Ordinary Refresh replaces inspected metadata with the basic inventory; read details again when needed. Selection and disclosure expansion perform no metadata requests. Reading details does not generate, load, unload or download models.

Compare… sends the selected model to the Hosted & mixed comparison setup without starting a run. Text models can be compared with ready native models. ComfyUI image comparisons require checkpoint models and an explicitly imported API workflow. Host inventories remain separate from local scan, conversion, verification and reclaim state.

Release host memory

The toolbar memory panel also lists saved hosts and their dated residency and activity readings. Open host selects the corresponding connection here; the refresh button beside each host reads only that host without loading or unloading models.

Ollama's memory summary includes its complete reported running list, even when installed tags differ. A successful empty running list reports zero; a failed or malformed list leaves residency unknown. Running-only weights do not become comparison candidates or unload targets. Use Ollama to manage weights absent from or different from the installed tags. Workbench unload still requires a loaded installed tag with the same digest.

Select a model reported as loaded, then Unload model… for Ollama or Unload instance… for LM Studio. LM Studio requires its native loaded-instance identities; when several instances are loaded, choose the exact instance first. Legacy inventory without those identities cannot authorize unloading. ComfyUI offers Free host memory…, which requests cleanup of all cached models on that host, independently of the selected checkpoint. llama.cpp and generic OpenAI-compatible connections do not offer a supported unload action.

The review shows the target, host address, activity and check time. ComfyUI reports running and queued jobs. Ollama and LM Studio remain Activity unknown because their public inventory APIs do not report other clients' active requests. Missing or failed activity checks never imply idle. Busy, queued and unknown activity carry a warning; Cancel or Escape sends no memory-release request.

Explicit confirmation checks the target and activity again. Changed activity or model details require renewed confirmation. Activity is a snapshot; another client can start work after any check. Workbench blocks memory operations during its active comparisons and blocks host edits while an operation is in flight.

The action uses the host's memory-management API. It never deletes model files, quits the application, clears queues or interrupts shared jobs. ComfyUI cleanup may be deferred until work finishes, and later jobs can reload models. Ollama and LM Studio unload requests may affect other clients using the selected model or instance.

Workbench refreshes inventory after a request and distinguishes acceptance from observed unloaded state. ComfyUI does not expose cleanup completion, so acceptance does not establish released memory. Failed readback retains the dated inventory and reports that the effect could not be verified. A connection failure after submission can leave the outcome unknown; requests are never automatically retried.