AI Chat
The model runs on this machine. Ask it questions, show it an image, hand it a document to read — the conversation never leaves your computer.
Chatting with the model
Launch one of the models you've downloaded and start talking to it like any AI assistant — except this time inference happens on your own graphics card.
- Token-by-token streaming — The response appears as it is generated, instead of landing all at once when the whole answer is finished.
- Thinking mode — Turn it on when you need deeper reasoning (off by default); once the answer is done, the reasoning is tucked into an expandable block.
- Long context — Context length depends on the model you choose, up to 128K; if you genuinely go past it, the app trims the history and tells you so.
Turn these meeting conclusions into three points, in a formal tone.
1. The next release prioritizes the offline experience; cloud features are deferred pending evaluation. 2. Engineering will produce a feasibility assessment and schedule within this week. 3. Marketing materials start only once the assessment results are in.
Image understanding
Drag an image into the conversation, or pick a file from your computer with the attach button, and an on-device vision model reads what's in it.
- Just drop it in — The whole chat surface is a drop zone; there's no upload box to find first.
- Needs a vision model — All three base chat models offered during first-run setup have vision capability; if you're on a text-only model, the app asks whether to switch first.
- Several at a time — Up to 20 MB per image, and up to 5 images per message.
What's happening in this photo?
It's a stretch of coastline: a child is running through the shallows toward the camera, kicking up spray. The shot is backlit, with low cloud and distant surf behind.
Document answers
Attach a PDF, Word file, plain text or spreadsheet and ask about it directly; the answer cites the passages it used, and you can open one to read the original.
- If it fits, it reads all of it — When the document fits in the context window, the whole thing goes to the model and the status reads "Reading the document".
- If it doesn't, it retrieves — Longer documents switch to retrieving the relevant passages for each turn, with the status "Retrieving relevant passages" — a document is never too long to get an answer.
- Honest refusals — A PDF with no text layer, such as a scan, is called out plainly, with a pointer to run it through the Document Recognition workstation first.
What does this document say about latency?
The document's position is that latency is set by your local hardware rather than by datacenter load — inference makes no network round-trip, and there is no queue at peak hours.
API access
While the app is running, the chat module doubles as a local inference service; any tool that supports a custom endpoint can connect to it directly.
- Two compatible endpoints — Both OpenAI-compatible and Anthropic-compatible endpoints are provided, so tools written for either API mostly work as-is.
- Localhost only — The service listens on 127.0.0.1 by default; letting other computers on your network use it is something you turn on yourself.
- No quotas — No token billing, no rate tiers — the ceiling is your hardware.
- Base URLhttp://127.0.0.1:2918/v1
- Modelas listed on the app's "API Configuration" page
- API Keyany string (the local endpoint doesn't check keys)
An Anthropic-compatible endpoint is provided as well; localhost only by default.
Got the machine? Four steps.
Install, agree, activate, download models — most of the time is just the download.