Fuses your local files, notes, and on-device corpus with an optional, clearly-toggled web layer. Every answer is cited; nothing leaves the phone unless you flip the web switch.
The capability manifest lists each model's vision, PDF, OCR, and tool-use flags. Routing picks the model that both handles the input type and fits your device's memory budget — quantization is preferred when the tier budget is tight.
ModelRouter.classify() takes a repoId, filename, tags, libraryName, and pipelineTag and returns a Classification with .module and .confidence.
A quick tour of MLX vs GGUF on Apple Silicon, how we rank quants, and why a 3B-Q4 often beats a 7B in day-to-day chat.
Mist searches your files, notes, and on-device corpus first, then augments with a web layer if — and only if — you opt in. Both streams merge into one answer.
Every claim traces to a source. Tap a citation chip to open the matching file, note, or page — and see exactly which passage Mist pulled from.
Offline-only by default. The web layer is a visible toggle on the search bar — you always know, at a glance, whether a query will leave your device.
Mist is Nimbus8's on-device search module. It answers questions from your own stuff first — files you've dropped in, notes you've taken, pages you've saved — and only reaches out to the web when you explicitly turn that layer on.
Every answer is cited. The model that writes the reply is the same local runtime that powers the rest of Nimbus8, so a query can use your installed chat or vision model without a round trip. Mist is currently in beta: the retrieval layer is stable, the web connector is narrow and deliberately boring.
Mist maintains two retrieval paths and reconciles them at answer time:
Both streams are merged into one context window, deduplicated by URL and passage hash, and passed to a summarization model. The answer is written once, with citation anchors numbered in the order they appear.
Each numbered chip in the answer corresponds to a source card below. Local sources are labelled L; web sources are labelled W. Tapping a chip scrolls to the matching card and highlights the passage the model pulled from.
If the model tries to write a claim without a supporting passage in context, Mist drops the sentence rather than hallucinate a citation. You'll see a shorter answer before you'll see a fake source — that's intentional.
Mist is offline by default. The web layer is a visible toggle on the search bar — not buried in Settings, not enabled-by-default after a silent update. When it's off, Mist makes zero network calls; the query, the embeddings, the answer, and the citations all stay on your device.
When the web toggle is on, a single outbound request fetches search results and their linked pages. Nimbus8 does not proxy these calls, does not log them, and does not retain any of the fetched content past the current answer. See the privacy policy for the full data flow.
Mist indexes what you give it — nothing more. You can add files in three ways:
Embeddings live in the app sandbox. Uninstalling Nimbus8 deletes them; there is no server-side copy because there is no server.
Mist uses two model slots: a small encoder for retrieval and a chat model for summarization. Both ship as device-verified picks in the Nimbus8 catalog.
You can swap either slot from Mist's Settings. The capability manifest filters the catalog so you only see models that fit your device.
Only if you turn the web toggle on. With the toggle off, Mist runs fully offline over your local corpus — airplane mode is a supported configuration.
Mist searches your stuff first. The web layer is an optional augment, not the primary source. If an answer exists in your own files, that's where it comes from — and that's where it's cited to.
Not yet, and not without your explicit scope. iOS sandboxes these out of reach by design, and Nimbus8 will not ship a backdoor. If a future version adds Mail or Messages, it will be opt-in per source with a visible indicator.
Every embedding, every indexed file copy, and every cached answer is deleted with the app. Because nothing ever left your device, there is nothing to clean up elsewhere.