Run AI on this computer with Local on this device

Local on this device is an AI source that runs PlexiiDesk's AI on your own computer, with a free AI model from PlexiiDesk that downloads once, when you first turn it on. Under the button, Settings describes it as Included with PlexiiDesk; requests stay on this computer when the model is on this computer. When the model still has to be downloaded, it says One-time download, with the size, then requests stay on this computer. On a computer with too little memory, it says Needs more memory than this computer has.

It works offline and costs nothing per use: no PlexiiDesk credits and no API key. The model is smaller than cloud AI, so it is good at some jobs and weaker at others. This guide explains what it runs, what it needs, and how it compares with the other AI sources.

Before you start

• Plan: every plan. You do not need an account, credits or a key to use it.
• Memory: your computer needs 8 GB of memory or more. On a computer with less, the model is not offered.
• Disk space: if the model has to be downloaded, you need 2.5 GB free for it, plus 1 GB to spare.
• Processor: on a Mac with an Intel processor the model runs on the processor without graphics acceleration, which is much slower than on Apple silicon. Settings says so on such a Mac.
• Who starts here. PlexiiDesk moves you to Local on this device only when the model is already on this computer and has passed its checksum. That applies to a new install, and to an update where your AI source was Automatic. A Mac installed from the disk image on plexiidesk.com has the model: PlexiiDesk checks it once in the background the first time it opens, then switches.
• Without the model, nothing switches. After an update from inside the app, or on Windows, your AI source stays as it was, and PlexiiDesk offers the download instead. See The first time you open 4.3.40.
• A source you chose stays. If you had chosen your own Anthropic key, your OpenAI key, Plexii AI credits without Automatic, or Local on this device, that has not changed.
• The choice belongs to this computer. Set it on each computer you use.

What runs on your computer

The model is Qwen3 4B Instruct, under the Apache-2.0 licence. A second, small model, nomic-embed-text v1.5, comes with it and lets PlexiiBrain search your work by meaning.

Settings describes the model this way: An AI model that runs on this computer. It works offline and costs nothing per use. It is smaller than cloud AI: good at answering from your desks and documents, writing, summaries and tidying; weaker at building whole desks and at agents.

Choose Local on this device

1. Click the gear at the right-hand end of the header to open Settings, then click the AI tab.
2. Under AI · source, click Local on this device. The chosen button is highlighted.
3. Under the buttons, a note says where your AI now runs. With no API key saved, it reads: PlexiiAssist, PlexiiBrain and the other AI features run on the on-device model below. No prompt or content is sent to any AI service; a feature the on-device model is not good enough for asks before using cloud AI for that one request.

If some features send their requests elsewhere, the note names each one and where it goes instead, and ends Switch any of them back to this device under Feature by feature.

Further down the same tab, the section headed Local on this device shows the model, how each feature runs, and the controls described below.

The first time you open 4.3.40

A card appears once, near the bottom of the window. Which card you see depends on whether the model is on this computer.

When the model is here, the card is headed Local AI is now built in. It says what the model is good at, and where your AI runs now.

• If your source moved to Local on this device, the card says Your AI now runs on this computer. Click Switch back to and the source you had, to go back.
• If your source stayed the same, the card says You are using and your source, then that has not changed. Click Use local AI to switch.
• Click Got it to close the card. It does not appear again.

When the model is not here, the card is headed Local AI is available as a download. It says: PlexiiDesk can run AI on this computer: it works offline, costs nothing per use, and sends nothing to any AI service. It needs a one-time download of, then the size. Under that, it says You are using and your source, then that has not changed. You can download it later in Settings › AI › Local on this device.

• Click Download, with the size beside it, to download the model now. A bar shows how far it has got, with Cancel to stop.
• Click Maybe later to close the card. It does not appear again. You can still download the model in Settings.

If your source was Automatic, the switch waits for the model. Once it has downloaded and passed its checksum, whether from the card or from Settings, your source moves to Local on this device and the Local AI is now built in card appears, with Switch back to. If you chose a different source in the meantime, PlexiiDesk keeps your choice.

Neither card appears while PlexiiDesk is still checking a model that came with it, or on a computer with too little memory to run the model.

The one-time download

Whether you need the download depends on how PlexiiDesk got onto your computer:

While AI is set to run on this computer and the model is not here, a banner near the top of the window says Local AI needs a one-time download, with the size. When Local on this device is your source, it goes on: AI is set to run on this computer, but the model is not here yet. Until it is downloaded, AI features cannot answer; or choose another AI source in Settings › AI. When only some features are routed to this device, it names them instead, and ends everything else is unaffected.

• The banner has Download and Settings buttons.
• To dismiss it, click × at its top right. It stays away until you next open PlexiiDesk, and comes back then if the model is still missing.
• It does not appear while the Local AI is available as a download card is showing, or while AI is working on another source.

If your source is not Local on this device, nothing changes while the model downloads. Under Local on this device in Settings, a note says Until then your AI keeps using the source chosen above; nothing changes until the download has finished and been checked.

To download it:

1. Make sure you are online.
2. Click Download in the banner or in the Local AI is available as a download card. You can also click Download in the model's card under Local on this device in Settings > AI.
3. Wait while it downloads. A bar shows how far it has got, for example 1.2 of 2.5 GB. The small search model downloads after it, if it is not already here.
4. At the end, PlexiiDesk checks the file against a checksum built into the app. Settings says Checking the file’s checksum… while it does.
5. When it is done, the model's card says Downloaded and verified, and the banner goes away.

The model downloads from PlexiiDesk's own server, dl.plexiidesk.com. Together with the search model it is about 2.6 GB, so the banner may show 2.6 GB while the card in Settings shows 2.5 GB for the main model alone.

You can keep working during the download. AI features answer once it has finished.

Stop or resume a download

• To stop, click Cancel in the model's card. What has arrived so far is kept.
• To carry on, click Download again. It resumes from where it stopped, whether you cancelled it or the connection dropped.

If the download does not finish

Check that it works

Under the model's card, click Try it. PlexiiDesk asks the model one short question on this computer and shows the answer, how long it took, and how long until the first words appeared. The first time includes loading the model.

A line under Try it shows the model's state:

• Starting local AI… this happens once per session. The model is loading. This takes a few seconds.
• Loaded and ready, with how long the last start took.
• Not loaded right now: it starts in the background after launch, or when first used, and unloads after 10 minutes unused.

While the model is loading and something is waiting for an answer, a small note near the top right of the window says Starting local AI — this happens once per session…

What runs here, and how well

In Settings > AI, under Local on this device, the list headed Feature by feature shows each group of AI features, where it runs right now (Now:), and how the on-device model scored against cloud AI in PlexiiDesk's tests of the app's real prompts.

What this means in practice:

• A feature that scored lower still runs on this computer, with a note the first time in a session: the feature's name, then ran on the on-device model. Works best with cloud AI; local results may be incomplete. The note links to Upgrade with your own Anthropic or OpenAI key.
• Images and background removal have no on-device model. Each request asks first before it goes to OpenAI or remove.bg.
• A request that needs tool use, an image or an attached document cannot run on the on-device model, so PlexiiDesk asks before sending that one request to cloud AI.
• If an answer comes back in a shape PlexiiDesk cannot use, it asks whether to try cloud AI for that one request.

Writing and drafts

While Local on this device is your source, a menu at the top of Feature by feature, Writing and drafts:, decides where writing, PlexiiWrite, PlexiiCells and PlexiiDeck generation, and Build with AI run:

If both keys are saved, a Prefer: menu beside it chooses Anthropic or OpenAI.

If a saved key is refused or out of credit, that request is answered on this computer instead, with a note such as Your Anthropic key was refused, so this was answered by the on-device AI. and a link to Check your API keys.

When a feature scored lower on this device

Two choices, under the same heading in Settings:

• Run everything this device can: weaker features run here with a label; if an answer comes back unusable, you are asked whether to try cloud AI for it. This is the default.
• Ask before using cloud: features that measured well run here; the rest ask before sending that request to cloud AI.

Either way, Settings says: Either way, nothing goes to cloud AI unless you say yes.

Route one feature somewhere else

Each row in Feature by feature has a menu that sets where that feature runs: Default, On this device, Cloud AI (account default), Your Anthropic key, Your OpenAI key or Plexii AI credits. A source you have not set up is marked (no key saved) or (sign in) and cannot be chosen.

• Test with local runs that feature on this computer until you switch it back. It appears on any feature that is not already running here, except images and background removal.
• Back to default returns the feature to your AI source.

Routing works from any AI source. You can keep Plexii AI credits as your source and still run, say, PlexiiAssist answers on this computer.

When PlexiiDesk asks to use cloud AI

When a request cannot run on this computer, a dialog asks Use cloud AI for this request? It explains why, says where the request would go, and says: AI is set to run on this device, so nothing has been sent. If you choose cloud, only this one request goes to cloud AI; everything else stays on the device.

• Use cloud for this request sends that one request. It takes the automatic path: your Plexii AI credits when you are signed in and they can serve it, otherwise your own Anthropic key.
• Keep it on this device sends nothing, and that request is not answered.
• Settings sends nothing and opens the AI settings.

If no cloud source is set up, the dialog says so, and Use cloud for this request cannot be clicked. If you do not answer within two minutes, nothing is sent.

What stays on this computer

While Local on this device is your source, PlexiiDesk refuses to send an AI request anywhere unless you said yes to it or set it up yourself. That covers requests to Anthropic, OpenAI, remove.bg and Plexii AI credits, and the web searches PlexiiAssist can make, because a search carries your question.

The exceptions are ones you choose:

• requests you approve with Use cloud for this request
• features you route to a cloud source under Feature by feature
• writing, documents and Build with AI while Writing and drafts is set to Use my own key when saved, else this device and a key is saved. Choose Always this device to keep these on your computer too.

Speech is transcribed on this computer too: while this is your source, voice notes and hold to talk use Local transcription, whatever the transcription setting says. The first transcription downloads the Whisper speech model from Hugging Face (huggingface.co) without asking and keeps it on this computer; your audio stays here. See Choose how speech is transcribed: Cloud or Local.

Things that are not AI requests carry on as usual: signing in, sync, downloading the model, and your own browsing.

See which AI answered

Many AI results in PlexiiDesk show a line such as Answered by: Local (Qwen3 4B), Answered by: Anthropic (your key), Answered by: OpenAI (your key) or Answered by: Plexii AI credits, so you can see which source wrote what you are reading.

Check details in drafts

When AI writes a draft for you, PlexiiDesk checks it for specifics your request did not give it, such as figures, prices, dates and names. If it finds any, a panel headed Check these details appears at the bottom left:

• Replaced with a placeholder for you to fill in: the clearest cases in a draft written on this computer, replaced with a placeholder.
• Check these before you use them: the rest, left as written for you to check.

Drafts from cloud AI are checked too, but never changed: anything found is listed under Check these before you use them:.

How it compares with the other AI sources

For credits, see AI credits and where your AI runs. For keys, see Add your own API keys.

PlexiiBrain and your documents

The same model describes your documents in the background, and the small search model lets PlexiiBrain search by meaning on this computer. Both are under PlexiiBrain on this device in the same section. See Enrich documents with a local AI model.

Search by meaning works passage by passage. PlexiiDesk splits each document into short passages and gives each one a search vector, so a document is ranked by the passage that best matches what you searched for, not by the document as a whole. A specific question can then find a long document that mentions the answer only once. In the box headed What PlexiiAssist can see, the line Passages ranked by meaning: counts how many passages are ready, for example 120 of 150. See Enrich documents with a local AI model.

That article also covers Ollama. Local on this device needs nothing else installed. Ollama is an optional extra that PlexiiDesk uses only to describe documents and search by meaning, and only while the built-in model is not downloaded.

Switch away from Local

1. Open Settings > AI.
2. Under AI · source, click Plexii AI credits, Your Anthropic key or Your OpenAI key.

If the Local AI needs a one-time download banner was showing, it goes away as soon as you switch, unless you have routed a feature to On this device. Set that feature back with Back to default to clear it.

The model stays on your computer. To free the disk space, click Delete in the model's card. Delete appears only on a model you downloaded: a model that came with PlexiiDesk shows Included with PlexiiDesk instead.

Tips

• Try it before you rely on it, especially on an Intel Mac, to see how long an answer takes.
• Want everything on this computer? Set Writing and drafts to Always this device and leave every feature on Default.
• Want better writing without giving up local answers? Save your own Anthropic or OpenAI key and leave Writing and drafts on its default. Writing, documents and Build with AI use the key, and everything else stays here.

Common questions

Does Local on this device use my credits? No. It costs nothing per use. Credits are used only for a request you send to Plexii AI credits.

Does it work offline? Yes, once the model is on your computer. Features that need cloud AI ask first, and cannot go ahead offline.

Why did my document draft use my Anthropic key? Writing and drafts is set to Use my own key when saved, else this device, which is the default. Choose Always this device to keep drafts here.

It says AI features cannot answer. The model is not on this computer yet. Click Download in the banner, or choose another AI source. To hide the banner for now, click ×.

I clicked Maybe later. How do I get the model now? Open Settings > AI, and click Download in the model's card under Local on this device. If your source was Automatic, it moves to Local on this device once the download has finished and passed its checksum.

It says "The on-device model is not downloaded yet." Open Settings > AI and download it under Local on this device, or choose another AI source.

It says the model needs 8 GB of memory. The card shows Needs 8 GB of memory; this computer has, and the amount your computer has. If you use an AI feature anyway, it says: This computer has your amount GB of memory and the on-device model needs at least 8 GB, so AI cannot run on this computer. To use AI, choose your own Anthropic or OpenAI key, or Plexii AI credits, as the AI source in Settings › AI. There is nothing to download that would fix it: choose another AI source on this computer.

Answers are slow. The first answer after launch includes loading the model, which takes a few seconds. On an Intel Mac every answer is much slower. Use Try it to see the timing, or route slower features to a cloud source under Feature by feature.

Can I use my own Ollama models instead? Not for PlexiiDesk's AI features. Ollama is used only for document descriptions and search by meaning, and only while the built-in model is not downloaded. See Enrich documents with a local AI model.

Related

• AI settings
• AI credits and where your AI runs
• Add your own API keys
• Enrich documents with a local AI model
• Choose how speech is transcribed: Cloud or Local
• Privacy settings and what leaves your computer