tensai

Theme

Language

Changelog

What changed in each version of the app, newest first. Also shown in the app: tap the version in Settings.

Atom feed

0.3.6#Latest

Faster replies, fewer stray tools

With thinking on, a simple question no longer makes you wait a minute, and chats with tools reply faster. Stories, poems and songs are written straight away instead of setting off searches.

Faster

  • With thinking on, a simple question gets a short think and a real problem a long one: a one-line answer now comes in seconds instead of about a minute
  • Chats with tools write their replies faster on the phone's processor
  • After a reply that used no tool, such as a haiku, the next question no longer waits about two minutes for its first word, even after closing and reopening the app

Writing

  • Asking for a story, poem, essay, song, rap or dialogue at the start of a chat, in English or French, gets it written directly, with no search or tool card
  • After two tool steps in a row fail or are refused, tensai answers with what it has instead of trying more tools

Instructions

  • Settings > Instructions lets you write instructions that apply to every chat; each chat can also have its own in its settings
  • Each chat now starts with a short built-in guide for the model, so an existing chat takes a little longer on its first message after the update

Time

  • Asking the time in another city, like Tokyo, gives that city's time instead of the phone's
  • Asking the time here no longer fails when the model guesses the time zone wrong

Chats

  • Swipe a chat left to delete it, with Undo for a few seconds

Fixed

  • A model import cut short no longer leaves a hidden file taking up storage; it is cleaned up at the next start
  • Switching between chats quickly or leaving the app can no longer mix up what the model remembers of each one

Privacy

  • Engine error messages no longer include your prompt or the reply text

0.3.5#

Web pages can no longer act behind your back

A web page tensai reads could try to make it send your data somewhere or plant a memory. Both now stop at a card you answer. Stop also works every time, even in a chat you left and came back to.

Web

  • Reading a page you named or a link a search returned works as before; any other address shows a card with the site and the full address first
  • So text from a page cannot quietly send your clipboard, contacts or notes to another site

Memory

  • Once a chat has read a web page, Wikipedia, your calendar, the clipboard or past chats, remembering a new fact asks first and shows the fact exactly as it will be stored
  • Asking tensai to remember something in a chat that read none of these works without a card

Privacy

  • The engine no longer writes the prompt to the system log

Stop

  • Tapping Stop right after Send now always stops the reply, instead of letting it run on unseen and keep the model busy
  • Leaving a chat while a tool runs and coming back shows Stop again; other chats wait for that reply instead of starting one that would collide with it
  • Stop in the notification ends the whole reply, tool steps included
  • A reply cut by an engine error is shown as failed instead of as a finished answer
  • Stop during the first instant of a reply no longer leaves the model's memory of the chat slightly off

Tools

  • When the model writes a tool call as raw text, tensai now runs the tool instead of showing the code as its answer

0.3.4#

Clearer confirmations, cleaner maths

Asking tensai to forget something now shows exactly which fact goes, and products like 17*23 no longer turn a reply into italics.

Memory

  • The card that asks before forgetting a fact shows the fact itself, not its number; a long fact is shown in full under the title
  • If the fact changed or disappeared between the card and your answer, nothing is deleted

Fixed

  • An asterisk between numbers, as in 17*23=391, stays an asterisk instead of turning the text around it into italics, in replies and in the reasoning block

0.3.3#

New app id, instant first message

tensai now installs as io.github.tensaiapp, and a new chat no longer waits for the model to read its tool list: the first message starts in seconds.

New app id

  • This version installs as a new app, next to the previous one: Android treats a new id as a different app, so it cannot update it in place
  • Models and conversations stay in the previous app: import or download your models again here, then uninstall the old app when you no longer need it

Faster

  • Right after a model loads, tensai reads the system prompt and the tool list in the background, so the first message of a chat starts in about 1 s instead of more than a minute (Qwen3-1.7B on a Dimensity 900, CPU)
  • Opening a new chat after changing the tools in Settings prepares the new tool list before you type, after saving the previous conversation: first message in about 4 s, and the previous conversation comes back intact

Under the hood

  • Vulkan can split large matrix products into shorter GPU jobs to keep the interface smooth while a prompt is read; off by default while it is measured on more phones

0.3.2#

Faster follow-ups

Follow-up messages no longer pay again for the whole previous answer, Auto picks the faster backend for more models, and Vulkan loads keep the phone usable while they read a prompt.

Faster

  • A new message reuses the cached previous answer instead of reading it again: after a 700-token reply, the next message starts in about 5 s instead of about 50 (Qwen3-1.7B on a Dimensity 900)
  • Auto now runs models of an unknown architecture on the CPU, which was faster everywhere it was measured on this class of phone (TwIL-LM3: prompt 14 tok/s on CPU against 3.9 on Vulkan)

Vulkan

  • Models that run on Vulkan read the prompt in smaller batches and on a low priority GPU queue: about 20% faster, and the interface no longer freezes for up to a second while the GPU works (it still runs below full frame rate)

Under the hood

  • Two tests that failed now and then in CI are deterministic

0.3.1#

Permissions + Network

The model can now read your calendar and your approximate location, and, if you turn it on, look things up on the web. The Internet tier is off by default: with it off, nothing leaves the phone.

Web (Settings > Tools > Internet, off by default)

  • Web search: DuckDuckGo, with Wikipedia and Marginalia as fallbacks, no key needed. Optional Tavily key (1,000 free searches a month, no card) or Brave key (monthly free credit, card required) in Settings, with a Test button
  • Read a web page: the model fetches a link and reads its text. HTTPS only, local and private addresses are blocked
  • Weather from Open-Meteo: today, tomorrow, a weekday or a date, for a city or "here"
  • Wikipedia summaries
  • Every network call shows where it went and what was sent: "Sent to DuckDuckGo · <query>"

Phone, with Android permission at first use

  • Calendar: "what is on my calendar today?", a day or a range up to 31 days. Only titles, times and places reach the model
  • Location: approximate only (about 1 km), never precise GPS. tensai does not turn it into an address
  • Revoking a permission while the model uses it gives a clear refusal with a link to the app settings

App

  • What's new: tap the version in Settings
  • Settings > Storage shows the real free space, with a colored legend (models, other data, free)

0.3.0#

Tooling

The model can now use tools: it checks things on the phone, acts through your apps, remembers what you tell it, and asks when your request is ambiguous. Everything still runs on the phone.

Tools

  • On this phone: date and time, calculator, unit and time zone conversion, device info (battery, storage, RAM), search past chats, memory, clipboard, torch
  • Phone actions through your apps: timer, alarm, calendar event, SMS and email drafts (to a number or a contact name), open an app or a link, media control. Nothing is sent by tensai: the target app opens and you confirm there
  • Contacts: drafts accept a name ("Jade"); with several matches the model asks you which one. Android's contacts permission is asked at first use
  • Questions: the model can ask a single or multiple choice question; the card replaces the input bar
  • Memory: facts you ask it to remember are added to every conversation; manage or turn them off in Settings > Memory

In the chat

  • Tool calls show as a discreet line between the reasoning and the answer; tap it for the timeline and each call's inputs and output
  • Calls that delete something or open a link ask first (Allow / Deny)
  • Settings > Tools: turn each tier on or off
  • Context sheet: tool definitions, tool results and memory in the breakdown, and an Agent activity section (calls, tokens, time per tool)
  • Copy a message: long press, or the copy button under replies

Faster

  • The fixed part of the prompt (system and tool definitions) is cached on disk: the first tool turn after loading a model went from ~4 min to ~10 s
  • CPU models no longer send the prefill to the GPU: prompt processing ~4x faster, generation ~1.7x faster on a Dimensity 900
  • The UI no longer stutters while a prompt is processed
  • The reasoning of the current turn is kept between a tool call and its result, so the model continues instead of starting over

0.2.0#

Models

Models from Hugging Face, a context view, and a simpler navigation.

Models

  • Discover: search Hugging Face, filters (fits my phone, tools), models recommended for your device
  • Repo page with every quant, its size and whether it fits; download with pause/resume, sha256 check, progress notification, mobile data and free space checks
  • Gated models with a Hugging Face token
  • Model picker from the chip in the input bar
  • Chat template fallback for GGUFs that ship without one
  • Files this build can't run (unknown architecture, draft heads) are flagged and can't be loaded
  • Per-architecture backend choice extended (lfm2 on CPU, nemotron_h on Vulkan)

Chat

  • Context ring in the top bar and a Context sheet: token breakdown, runtime details
  • The app opens on a new chat; the old Home screen is its empty state
  • Chats page: search, grouped by date, multi-select delete
  • Input bar capped at 10 lines, scrolling past it

Navigation

  • One conversation at the root; Models, Settings and Chats open on top and back returns to it
  • Drawer scrolls as one list, recents capped at 20 with "All chats"

0.1.0#

First release: run a GGUF model fully on your phone and chat with it. Nothing leaves the device.

Models

  • Import a GGUF file from this phone
  • llama.cpp engine on CPU and Vulkan, backend picked per architecture
  • Session cache: a conversation you reopen resumes without re-reading its history

Chat

  • Markdown replies with code blocks, and the model's reasoning in a collapsible block
  • Per-conversation settings: instructions, thinking on or off, sampling presets, max response length
  • Onboarding on first launch, Home screen, empty and error states

App

  • Settings: theme, language (English and French), backend, context length, KV cache type
  • Signed release APK built by CI