About Brello

Our mission

Our mission is to build capable intelligence that answers only to the person using it, with privacy that can be verified rather than promised.

Using a capable assistant usually means sending your questions, photos and answers to someone else’s computers, where a policy decides what is kept. We treat privacy as an engineering constraint instead. Brello 1.0 runs its AI entirely on the phone, and Brello Super Intelligence is being designed so that work done off the device is not retained and can be checked by independent researchers.

The reasoning behind the mission, and the commitments that follow from it, are set out in the Brello Charter, version 1.0.

What Brello Research publishes

Brello Research is the name under which our team publishes its work: engineering notes and research, safety and design papers about Brello 1.0 and Brello Super Intelligence. Publications are credited to the team, not to individuals.

Notes about Brello 1.0 describe the app as shipped, with the parameters it actually uses. Notes about Brello Super Intelligence describe design intent, not results, and say so. Every publication carries a dated version history.

The work covers six areas: efficient on-device inference, verifiable private compute, grounded answers, perception, personal context, and alignment and evaluation. A good place to start is “Private compute you can verify: the design space”, or browse every Brello Research publication.

How we work

Five practices shape the work. Each one is visible in Brello 1.0 today, or written into the Brello Charter for Brello Super Intelligence.

1. 01

   Build privacy into the system, not the policy.

   Where a privacy property can be enforced by design, we design it in. Brello 1.0 runs its models on the phone, has no account system and no Brello server, and keeps chats only in its private app storage, excluded from Android cloud backups and device transfers. If you turn on web search, the search text goes directly from the phone to a search engine, which sees an ordinary web request, including your IP address.

2. 02

   Build the strictest version first.

   Before designing anything that runs off the device, we built an assistant with no server of its own. Brello 1.0 downloads one model once, between 977 MB and 3.66 GB, then answers offline on the phone’s own chip, using the GPU when it can and the CPU otherwise. Brello Pro and Brello Vision understand photos offline too.

3. 03

   Publish the mechanism, with its numbers.

   We document how Brello works with the parameters the app actually uses. When Brello 1.0 answers from the web, it reads up to four result pages, splits them into passages of about 420 to 700 characters and ranks them with BM25 (k1 = 1.2, b = 0.75), as set out in “Answering from the open web, without a server”.

4. 04

   Evaluate before shipping.

   Automated tests check Brello 1.0’s model recommendation, repetition guard, search heuristics, query refinement and relevance guard, and a live end-to-end test exercises its web research. For Brello Super Intelligence, we will test how each new capability could fail or be misused before it ships, and publish what we evaluate and what we find. The method is in “Evaluate first, then ship”.

5. 05

   State each limit beside the claim it qualifies.

   Every Brello model names the open model it is built on, and our pages give download sizes, memory needs and limitations in the same place as the claims they qualify. Brello 1.0 runs only on Android, takes one photo per message, has no voice input and carries only the last six messages into each answer.

Milestones and status

Brello 1.0 was built on 4 October 2026, and we published the Brello Charter and our first research on 5 October 2026. Brello Super Intelligence is in development.

As of 5 October 2026, Brello Super Intelligence is in development. It will open to a small group first, and we have not announced a date. Our announcements are listed by date in Brello news, starting with “Introducing Brello Super Intelligence”.

Careers and open applications

We have no roles listed, and we welcome open applications. If you work on reasoning, on-device inference, privacy engineering, safety evaluations or interface design, write to us, even if no role is listed.

Tell us what you have built or published, and what you would like to work on. Each link below opens your email app with a prepared subject line for that area.

• ReasoningModel behaviour, reasoning modes and grounded answers
• On-device inferenceRuntimes, memory and mobile GPUs
• Privacy engineeringHardware isolation, remote attestation and verifiable systems
• Safety evaluationsRed-teaming, measurement and release gates
• Interface designInteraction, motion and writing

Contacts: general, press, security, privacy

Write to us by email. The general, press and privacy links open your email app with a prepared subject line, and security reports follow our disclosure policy.

All four go to [email protected] Address to be confirmed. Journalists will find a fact sheet, naming guidance and screenshots in the Brello press kit. For how Brello 1.0 handles your questions, photos and answers, read privacy at Brello and the privacy policy.

Naming and attribution

Brello is one word with a capital B. These are the names we use and what each one refers to.

Brello’s models are built on open models from Google and Alibaba, released under the Apache 2.0 licence. Brello is not affiliated with or endorsed by Google or Alibaba. The naming guide in the press kit covers first mentions and the use of images.

Editorial standards

Four rules govern what we publish. Our editorial standards set them out in full, with how we handle references, screenshots and bylines.

• Facts about Brello 1.0 come from the app itself, version 1.0.0, in the app’s units. Sizes are decimal, so 1 GB is 1,000,000,000 bytes.
• Brello Super Intelligence is described as design intent, with no claimed results, release dates, partners or customers.
• Comparisons with other assistants quote each provider’s own published page, with a link and the date we checked it.
• Research and news pages carry a dated version history, and material corrections are recorded there rather than made silently.