Now in early access · Android
New Read PDFs, photos & shared chatsRuns open-weight models fully on-device. Works offline. Nothing leaves your phone unless you ask it to.
Download APK — coming soon Android 9+, 64-bit
Animated preview of the BentoMind chat: reading an attached PDF offline, summarizing a chat shared from WhatsApp, an offline question, a nearby places search using OpenStreetMap, and a calculator tool call, each with live speed, CPU, memory and temperature readings.
Powered by open models via llama.cpp
New · Attach & share
Attach files from the + button, or share them from any app. Documents are read on your phone, so your lease never leaves it.
BentoMind pulls the text out of a PDF, a .txt, a CSV or code right on the device, so even an offline model can summarize a lease, a bill or a report.
Offline · any model “What's the notice period, and when is rent due?”
Tap Share → Ask BentoMind in WhatsApp, Gallery, Files or Chrome, or select text anywhere. Then one tap: Summarize, Draft a reply, Dates & tasks, Explain or Translate.
No SMS or storage permission “Summarize this group chat. Anything for me?”
Ask about a screenshot, a menu or a whiteboard with a vision model you connect (Ollama, LM Studio or any OpenAI-compatible API). On-device models say honestly when they can't see.
Remote vision model “What does this sign say in English?”
Features
Everything you expect from a modern AI chat, plus the one thing cloud assistants can't offer: it works with your phone's data without that data ever leaving it.
"Am I free at 4?" "Pharmacy near me." "Add Rahul's new number." "Summarise this chat." BentoMind uses typed tools for contacts, calendar, location and notifications, and you can share any text or chat to it from other apps.
No account, no server, no analytics. What you ask stays on your phone.
Honest performance: live tokens/s, CPU, RAM and temperature while you chat, plus a benchmark lab to find the model that suits your phone. Export results as CSV or JSON.
On a flight, in the mountains, or on a patchy connection. Once a model is downloaded, chat, local voice input and on-phone tools like contacts and calendar work in airplane mode. No per-query cost, ever.
Speak instead of typing. Use Google's on-device recognizer, or a fully local Moonshine or Whisper model.
4 s of audio → text in 83 ms
Auto thinking reasons only when it helps (math, code, planning). A budget and loop detection stop small models from going in circles.
Ask it to remember something and it saves a short, editable note on your phone. View, edit or delete any memory at any time.
Pick from 20 open-weight models sized for your phone, from 292 MB to 3.1 GB, with a device-fit badge on each. Paste any GGUF link from Hugging Face. Or connect an OpenAI-compatible server you trust, such as Ollama on your laptop; API keys are encrypted on the device.
Use cases
Real things to ask today. Most of them work in airplane mode, and every card says what it needs.
“What did I miss? Just the plans, and anything for me.”
Export or share the chat → Summarize
Offline
“Help me say no to Saturday without sounding rude.”
Share the message → Draft a reply
Offline
“What is this charge, and when is it due?”
Select the text in any app → Ask BentoMind
Offline
“Uh, call the plumber, Mum's birthday is Sunday, and renew the passport…”
Tap the mic and think out loud
Offline · Voice
“What's the notice period, and when is rent due?”
Attach the PDF → the text is read on your phone
Offline · PDF
“What does this sign say in English?”
Attach or share a photo, with a vision model you connect
Needs internet · Remote vision model
“Five quick dinners for this week, and one grocery list.”
Then ask it to swap anything you don't like
Offline
“How do I say “Where is the train station?” in Japanese?”
Abroad or in airplane mode, no roaming needed
Offline
“Explain photosynthesis like I'm 12, then quiz me.”
Ask follow-ups until it clicks
Offline
“Am I free at 4 tomorrow?”
Checks your calendar, right on the phone
On-device data · Calendar
“Who messaged me while I was in the meeting?”
Looks through your notifications, on the phone
On-device data · Notifications
“Is there a pharmacy near me?”
Searches OpenStreetMap with an area rounded to about 100 m
Needs internet · Location
“What's the weather today, and the top headlines?”
Web search via DuckDuckGo, with sources
Needs internet · Web search
“Remember I'm vegetarian.”
Later recipe ideas skip the meat. Edit or delete the memory any time.
Offline · Memory
How it works
One tap. BentoMind reads your phone's RAM and recommends models that will run well, with a fit badge on every option.
from 292 MB · resumable download
Type or speak. Answers stream token by token with Markdown, code, tables and LaTeX, and your history stays on the phone.
loads in 1.5–3 s · works offline
Switch on only the sources you want: contacts, calendar, location, notifications. Every change it proposes waits for your OK.
opt-in per source · OK / Cancel
Privacy
The model runs on your phone's CPU, so your questions, chats and personal data are processed in your pocket. Data only crosses the line for things that need the internet, and you control each one.
No account, no server
There is nothing to sign up for and no BentoMind backend that sees your chats.
No ads, analytics or trackers
No tracking SDKs in the app. Your data is never the business model.
Only the minimum, under your control
Web search sends just the query and can be switched off. Nearby places sends an area rounded to about 100 m.
Delete anything
Clear chats, memories and models in the app. Uninstalling removes everything.
on-device boundary
llama.cpp
on your CPU
Web search
Query only → DuckDuckGo, when online. Switch it off any time.
Remote model
Opt-in. Only to a server you add, labelled in chat.
Models & performance
No cherry-picked flagship. Every figure below was measured in BentoMind's benchmark lab on an iQOO Z10 (Snapdragon 7s Gen 3, 8 GB RAM).
Decode speed, higher is better Prompt processing (pp512), higher is better Time to first token, lower is better
4 threads, Q4_K_M / Q8_0 GGUF, measured 4 Oct 2026. *Qwen3.5 2B measured in chat with a 692-token prompt. Decode speed is limited by memory bandwidth, so smaller files are faster.
1.5–3 s
to load a model
83 ms
to transcribe 4 s of speech (Moonshine)
Need more power?
Point BentoMind at your own GPU. Qwen3.5 4B via Ollama on a GTX 1650 laptop streams at ~38–44 tok/s over Wi-Fi.
20 models in the catalog 292 MB – 3.1 GB
| Model | Quant | File | Generation tok/s | Prompt tok/s | First token | Peak RAM |
|---|---|---|---|---|---|---|
| MiniCPM5 1B | Q4_K_M | 688 MB | 13.8–14.4 | 112–118 | 552 ms | 1.40 GB |
| LFM2.5 1.2B Instruct | Q4_K_M | 731 MB | 13.5–14.4 | 101–125 | 541–622 ms | 1.72 GB |
| Qwen3.5 0.8B | Q8_0 | 812 MB | 10.5 | 131 | 422 ms | — |
| MiniCPM5 2B | Q4_K_M | 1.6 GB | 6.4–7.0 | 39–50 | 1128 ms | — |
| Qwen3.5 2B | Q4_K_M | 1.3 GB | 3.2* | — | — | 1.2 GB |
The app
Real screenshots from BentoMind running on an iQOO Z10.
Read PDFs · New
Attach a PDF; its text is read on the phone, even offline.
PDF answers · New
Rent, deposit, dates and terms pulled out by a 1B model.
Attach & share · New
Photos and files from the + button, or share from any app.
Chat
A clean start, with live metrics up top.
Code & Markdown
Streaming answers with syntax-highlighted code.
Web search
Searches DuckDuckGo when online and cites sources.
Auto thinking
Reasons when it helps, and stops overthinking.
Live metrics
Speed, first-token time, CPU and RAM per reply.
Model catalog
20 models with a device-fit badge for your phone.
Voice models
On-device speech recognition, from 28.5 MB.
Remote models
Use your own GPU server, clearly labelled.
Permissions
Switch on only the data you want it to use.
Settings
Theme, default model, system prompt and more.
BentoMind is in early access on Android. Tell us your phone model and we'll send you a build.