Brello Super Intelligence
Brello Super Intelligence (Brello SI) is the private AI assistant Stuvio is designing, and it is in development. It is being designed for deep reasoning, long-term personal context and tasks done on your behalf: running on your device whenever it can, sending heavier work only to sealed compute your device has verified, and using the web only when you allow it. Section 4 of the Brello Charter describes the three layers.
Not yet. Brello SI is in development, we haven’t announced a date, and it will open to a small group first, by invitation. We will announce the date in News.
By design, only when a task needs more than your device can give. Brello SI is being designed to send that work to sealed compute: hardware-isolated, stateless servers intended to keep nothing, which your device must verify before it sends them anything. We will publish the architecture before anyone outside the team uses it (Brello Charter v1.0, section 4). ‘Private compute you can verify’ sets out the design options.
No. ‘Brello Super Intelligence’ is a product name, not a claim about capability, and the assistant it names is in development with no published results. Before Brello can do something new, we will test how it could fail or be misused, and publish what we tested, how, what we found and what it still gets wrong (Brello Charter v1.0, commitment 04).
We haven’t announced pricing. We have committed not to show advertising in Brello or on this website, and not to use advertising to pay for Brello (Brello Charter v1.0, commitment 02). Brello 1.0 is a prototype with no public download, and Brello Super Intelligence is in development.
Brello 1.0: privacy and data
No. Brello 1.0’s language model runs on your phone, so your questions, photos and answers are processed there and never sent to Stuvio. If you turn on web search, the search text goes directly from your phone to a search engine, and up to four result pages are requested from their sites; the engine and those sites see an ordinary request, including your IP address. Privacy at Brello covers every data flow.
No. We will not use your questions, photos, answers or chats to train or fine-tune any model, or for anything except answering you (Brello Charter v1.0, commitment 01). Brello 1.0 never sends them to us: it processes and stores them only on your phone, and excludes them from Android cloud backups.
No. Brello 1.0 has no account system, sign-up or login, and its model files download from Hugging Face without an account or token. Because there is no Brello server, Stuvio holds no profile of you. The Brello 1.0 overview lists what you need instead.
Only on your phone, and they are not backed up. Chats are kept in a single file in Brello’s private app storage, excluded from Android cloud backups and from device-to-device transfer. Swipe left on a chat to delete it, or use ‘Delete all chats’ in Settings. With no sync or export, chats are gone if you lose or reset the phone (privacy policy, section 5).
Only the search text, and only when Brello searches the web. That is a cleaned-up version of your question, sometimes with up to ten words of your previous one, sent from your phone straight to DuckDuckGo, Bing, Brave Search, Google News or Wikipedia with ‘Do Not Track’ and ‘Global Privacy Control’ signals. The engine and the sites Brello reads see an ordinary request, including your IP address. Privacy at Brello details each request.
Internet and network-state access, for model downloads and web search, plus three permissions that keep a download running in the background: foreground service, notifications and wake lock. Camera and photos are reached through Android’s system picker only when you choose them. Brello doesn’t request contacts, location, microphone, SMS, calendar or storage-wide access. The Brello 1.0 system card lists each permission and why it is needed.
No. Brello 1.0 contains no analytics, crash-reporting or advertising software, so Stuvio receives no usage data from the app. The trade-off is that we learn about a crash or a wrong answer only if you tell us. The Brello 1.0 system card lists every data flow the app has.
Brello 1.0: using it
Brello 1.0 is a private AI assistant for Android and iPhone, made by Stuvio, that runs its language model entirely on the phone. You can ask it questions, ask about a photo and, if you choose, let it search the web and cite its sources. Version 1.0.0 is dated 4 October 2026. Read the Brello 1.0 overview.
Yes, after a one-time model download of 977 MB to 3.66 GB, depending on the model. Questions, photo questions and Think harder all run with no connection. Only web search needs the internet, and it is off by default. On-device AI, explained covers how this works.
Not unless you have turned web search on. With it off, a time-sensitive question, about news, prices, weather or scores for example, brings up a card titled ‘Search the web for this?’ with two buttons, ‘Search the web’ and ‘Answer offline’. Choosing to search turns web search on until you switch it off. See ‘Asking before going online’.
Your phone does the searching itself, trying up to eight search providers in turn, the first two through a private, invisible browser. Brello reads up to four result pages, splits them into passages of about 420–700 characters, ranks the passages with BM25 and keeps at most two per source. The model on your phone then writes the answer and cites its sources inline as numbered links. ‘Answering from the open web, without a server’ describes each step.
Yes, with Brello Pro or Brello Vision; Brello Core reads text only. Attach one photo per message from the camera or your gallery, then ask about it or send it alone for a description. Photos are processed on the phone and never trigger a web search. PDFs and other files aren’t supported. Brello Vision is the lighter photo model.
Think harder is a reasoning mode in which the model works through a problem before it answers, showing its reasoning in a ‘Thought process’ panel you can expand afterwards. It is slower, allows answers of up to 2,048 tokens instead of 1,200, and works with all three models. Reasoning modes, explained covers the trade-offs.
The last six messages. Each answer uses the six most recent messages as context, trimmed to 300 characters for yours and 600 for Brello’s, so early details in a long chat drop out; restate anything important. Brello has no memory across chats. Context windows, explained describes the limit.
Yes. Brello’s models are far smaller than frontier cloud models: they can be wrong, have a knowledge cutoff and are weaker at long or complex reasoning. The model is told, ‘If you are not sure about something, say so instead of guessing’, which is no guarantee, and answers from the web cite numbered sources. See Safety at Brello and why AI makes things up.
It has no voice input or output, takes one photo per message and no other files, and has an English-only interface. It has no sync, backup or export, runs only on 64-bit ARM Android phones, and its speed and quality depend on the phone. When web search comes back empty, it answers from its own knowledge. The Brello 1.0 overview lists its limitations.
From Google Play for Android and the App Store for iPhone.
Models, phones and storage
Android phones with a 64-bit ARM processor (arm64-v8a) running Android 8.0 or newer; virtually all modern Android phones qualify, but x86 devices and Chromebooks don’t. The smallest model is recommended for 4 GB of RAM. A GPU is optional: Brello uses it through OpenCL when it can and otherwise runs on the CPU. The Brello 1.0 overview lists the full requirements.
Yes. Brello 1.0 is on the App Store for iPhone and on Google Play for Android. The Brello 1.0 overview lists the Android phones it runs on.
The one Brello marks ‘Best’. Brello reads your phone’s memory and recommends the most capable model that fits, allowing 10% because phones report a little less than advertised: Brello Pro for 12 GB of RAM, Brello Vision for 6 to 8 GB and Brello Core for 4 GB. It warns you if another choice may be too big. See ‘Fitting a model to the phone in your pocket’.
Open models from Google and Alibaba. Brello Pro, the most capable, is based on Gemma 4 E4B; Brello Vision, which also understands photos, on Gemma 4 E2B; and Brello Core, the fastest and text-only, on Qwen3 1.7B. All three are released under Apache 2.0 and run with Google’s LiteRT-LM. Brello isn’t affiliated with or endorsed by Google or Alibaba (licences).
About 1.95 GB for Brello Core, 3.60 GB for Brello Vision and 6.26 GB for Brello Pro: the one-time download (977 MB, 2.59 GB or 3.66 GB) plus a speed cache built on first launch. Before downloading, Brello checks for that space plus 300 MB of headroom, about 2.25 GB, 3.90 GB or 6.56 GB free. Compare the three models.
Because Brello tunes the model to your phone’s chip. On first load, the runtime builds a cache of the model’s weights for that chip, the setup step labelled ‘Optimize for this phone’. It happens once per model and takes up to about a minute. Running a language model on an Android phone explains the steps.
Company, research and security
Stuvio, a small team building private AI. Stuvio makes Brello 1.0 and is designing Brello Super Intelligence, and publishes its commitments for both in the Brello Charter. Brello is built on open models from Google and Alibaba, neither of which endorses it. Read about Stuvio.
In Brello Research’s publications. The research index describes Brello 1.0 as built and sets out design notes for Brello SI. The Brello 1.0 system card collects its models, data flows, safeguards and limits, Learn explains the underlying ideas, from on-device AI to web retrieval, and the glossary defines the terms.
Brello 1.0 tells its model to say when it isn’t sure, cites web sources inline, asks before going online, stops answers that start repeating themselves and recovers when a model fails to start. Evaluations and release gates for Brello SI are planned, not done. Safety at Brello lists each safeguard, its mechanism and its status.
Email us with ‘Security report’ in the subject line, as our responsible disclosure policy describes. It sets out what is in scope, what to include and our safe harbour: if you act in good faith and follow it, we won’t pursue or support legal action against you for your research. Please allow reasonable time, normally 90 days, before disclosing.
No. This website sets no cookies, stores nothing in your browser and runs no analytics, advertising or third-party scripts; its fonts, images and scripts are served from this site. Like any website, its host receives standard request information, such as your IP address. Section 3 of the privacy policy has the details.