How the test went
The listing
The PromptQuorum directory lists Caelum as a free, closed-source Android chat app that says it runs AI offline, with document questions and optional web search. It records no named model, a website that no longer resolves, 10K+ Play downloads, Android 7.0+ and no update since 14 August 2025 (data read 19 October 2026; the update date is older than one year).
First launch and download (10:36 to 10:39)
The first screen reads 'downloading needed data...', 41 Mo/1020 Mo, ETA 3m 8s, with a tip box and 'Tap to see other tips...'. The 5.60 MB/s figure is the status bar's own indicator.
The app never says which LLM it is downloading: the screen shows only 'downloading needed data...'.
The next screenshot, at 10:39, shows the chat home screen. The end of the download was not screenshotted, so 'about three minutes' is derived from the two clocks.
The tester called the screen 'very colorful, very beautiful' and the input area 'a little bit crowded'.
The unit shown is 'Mo', the French abbreviation for megabyte, while the rest of the visible text is English. The tester calls the download 'one gigabyte'; 1,020 Mo is about 1.0 GB.
Home screen (10:39)
'How can I help you today?', two suggested prompts ('What are the benefits of meditation?', 'Explain quantum physics in simple terms'), and an input row: text field, blue globe button, green folder button, yellow send arrow.
The folder button is how the photo was attached; the notes do not say.
Photo input (10:43)
The tester sent a photographed DHL slip with the prompt 'what is the picture about'. The answer: a DHL package, weight 4.7 kg, a tracking number, 'sent from Germany to Switzerland'.
He was impressed that such a small app reads pictures, and wrote: 'Switzerland, not sure where it got that idea.'
Whether the weight, tracking number and route match the slip is not recorded: the slip itself is not in the notes. Only the destination was questioned.
The chat title reads 'where is paris?' while the visible message is 'what is the picture about'. The title is the chat's first question: the tester had added two further questions, including the photo question, to the same chat.
Action buttons: edit, regenerate and delete under the user message; copy and read-aloud under the answer. The screen was in dark theme at 10:43 and in light theme at 10:39; how the theme changed is not recorded. The tester: 'it has a dark theme. Nice.'
Web search (10:43 to 10:47)
While a rainbow-bordered box says 'Searching the web...', the globe button is darker blue and the send arrow becomes a yellow square. At rest the globe is lighter blue.
He found it totally unclear when web search is on or off and what it does: 'two different shades of blue'.
The globe is also darker blue in the Rwanda screenshot at 10:47, which suggests web search stayed on.
The Rwanda question (10:47)
In a new chat titled 'what is the politi...' the visible part of the answer says Rwanda 'has been experiencing significant political instability following the assassination of President Paul Kagame in July 2023', citing 'reputable news sources (like Reuters and BBC)'. No links are visible in the screenshot. The full question was not captured; it is assumed to be a question about the political situation in Rwanda.
A Google AI Overview, which the tester captured, says Kagame has been in office since 22 April 2000 and won re-election in July 2024; Google's own note says AI responses may include mistakes. The Overview shows a photo of a real person, so the screenshot is not published here.
A further web search by PromptQuorum on 21 October 2026 found his official site and an election report, and no report of an assassination. This is a limited check.
Logic stress test and cross-check (10:49 onward)
ChatGPT produced a prompt titled 'Advanced Logic Stress Test: The Blackout Protocol', described as testing constraints, temporal reasoning, contradictions and multi-step deduction. The tester pasted it into Caelum.
Caelum wrote 'a big report' and a minute later was still writing; it ended by offering a deeper dive. The text of its answer was not captured; the only visible fragment, in a pasted preview, begins '1. Feasibility Ve...' and 'The plan is highly improbable to...'.
The tester pasted the answer into ChatGPT. Its verdict: 'The AI sounds analytical, but it makes several serious logical errors. I would score this response quite poorly', with the 'correct-looking conclusion β no valid plan β without establishing the actual contradiction'. Its scores: 5/25 overall (constraint comprehension 2/5, formal deduction 1/5, temporal reasoning 1/5, proof quality 0/5, instruction following 1/5), 'not a standardized benchmark score'. ChatGPT also wrote: 'One important correction to my own earlier analysis'.
To test whether the prompt itself was unfair, the tester ran it through Grok (free version), Claude Opus 5.5 (medium) and Claude Fable 5.1 (high) and showed each answer to ChatGPT. ChatGPT rated the later answers better ('substantially better', 8/10 on logical deduction in one case) and still found errors in each, for example a count of 12 valid schedules that 'are actually 13', and a minimum-repair proof 'not yet rigorous enough'.
Opus answered 'in less than one minute'; Grok was still working after about three minutes. Neither was timed.
His rule: always cross-check, never trust a single model, 'especially not the smaller one'. He also says Fable was the first to explain what the task was.
The decision
The tester sums up: reading pictures is 'incredible' for a small model; web search is unusual among comparable apps (not checked by PromptQuorum); quality is mixed, which is no surprise from a 1 GB model. The local nature of the AI 'is actually not stressed at all', and he doubts a user would know the model runs locally.
Would he keep it? 'No, I don't see the point.' Free services on frontier models do much more, so the file is not worth the space.
Phone clock and battery
10:36: first screen, 41 Mo of 1,020 Mo downloaded, ETA 3 m 8 s, 5.60 MB/s, 5G, battery 70%.
10:39: chat home screen, light theme, Wi-Fi icon, battery 67%.
10:43: photo question, dark theme, web search starts, battery 66% and 65%.
10:47: Rwanda answer, battery 64%.
10:49: ChatGPT writes the logic prompt (in the ChatGPT app), battery 63%. 10:55 to 10:56: ChatGPT judges Caelum's logic answer (in the ChatGPT app), battery 59%.
Battery fell from 70% to 59% in 20 minutes with the screen on and with the ChatGPT app in use for part of that time, so none of it can be assigned to Caelum. Between 10:39 and 10:47, mostly inside Caelum, it fell from 67% to 64%. No memory chart is possible because the app's memory use was not recorded (the phone has 12 GB). Throughput readings in the status bar (0.03 to 0.58 KB/s in the chat screenshots) are single snapshots, not a traffic log, and do not show whether the app is offline.