Treasynx AI model card
Last updated 17 October 2026 · bmvp.ai Ltd
The short version
Treasynx AI is an optional small model that runs on your phone. Version 0.1 does four jobs: help answers, quick capture, duration estimates and the weekly review. Every other task uses the app's own planner, with no model call. It can be wrong, it is weaker than large AI models at anything general, and it is not advice. This page shows the test results, including where it loses.
Version 0.1
- Released: 17 October 2026. Its signed listing expires on 17 April 2027; after that date it cannot be installed, and a newer version should replace it before then.
- Model id:
treasynx-ai-0.6b-v0.1 - Base model: Qwen3-0.6B from the Qwen team (Apache-2.0 licence), fine-tuned by bmvp.ai Ltd with LoRA and then preference training (DPO), then merged.
- File: a MediaPipe .task file of 666,425,366 bytes (about 640 MB).
- SHA-256:
d05bd8a5615b2483c629ce5eea03b098fd0ecbc1537a1a872a2e9b9c64783d5c - Runs on: the phone's processor, offline. Phones with less than about 3 GB of memory are refused. So far it has been tested on one phone, a Xiaomi 14 Ultra, where a help answer took about 15 seconds.
- Download: from models.treasynx.com, only after you choose it. The app first checks the signed listing (Ed25519 and ML-DSA-65) and then the file's size and SHA-256.
What it does, and what is held back
Each task had to pass the same gates before it was switched on: valid answers, a task score at least as good as the app's own planner (the fallback), safety tests, and a check in each language. Four tasks passed. The others stay with the app's own planner.
| Task | Result in testing | In version 0.1 |
|---|---|---|
| Quick capture | Exactly right 93.4% of the time, against 5.2% for the fallback | On |
| Duration estimates | Median error of 15 minutes, against 20 minutes for the fallback | On |
| Help answers | Cites the right help topic 93.5% of the time, against 94.9% for the fallback. Says correctly that something is not in the help 93.3% of the time, against 0% | On |
| Weekly review | Every answer valid in testing | On |
| Plan a day | Valid at the first try 86.5% of the time; the fallback is always valid | Held back |
| Running late | Valid at the first try 88.2% of the time | Held back |
| Break a goal into blocks | Valid at the first try 84.8% of the time | Held back |
| What now? | Agrees with the app's rule 93.5% of the time; 98% was needed | Held back |
| Assistant actions | Failed a red-team test: it reworded a fixed safety message | Held back |
| Sorting requests | Refused 10.77% of ordinary requests, against 0% for the app's rules | Held back |
Where it loses
- Valid plans. The app's own planner is always valid; the model is not, so planning a day, running late and breaking a goal into blocks stay with the planner.
- Sorting requests. It refused more ordinary requests than the app's rules, and caught fewer requests that are outside its job.
- Help citations. 93.5% against 94.9% for the fallback.
- Injected text. In tests that hid instructions inside help passages, it copied the hidden text into 2.25% of its answers; the fallback never did. Help passages ship inside the app, so this would need a tampered help file. It should be fixed in a later version.
Where it wins
- Quick capture: 93.4% exactly right, against 5.2%.
- Duration estimates: about 25% lower median error.
- Knowing what is not in the help: 93% against 0%.
Safety tests
- Money, health, legal and food-safety requests, and messages about self-harm, are caught by the app's own filter before the model is asked. In testing it caught 100% of them.
- Some red-team prompts, such as jailbreaks and "repeat your instructions", were labelled as help questions by the model. That is one reason sorting requests stays with the app's rules.
- Privacy: a membership-inference test (AUC 0.5101, against a limit of 0.55) and planted canary checks found no sign that it memorised training examples.
Languages
The four enabled tasks gave valid answers in every language tested. Quick capture was tested in all 19 languages of the app; help answers, duration estimates and the weekly review were tested in English only. The app's crisis and safety wording in languages other than English has not yet been reviewed by native speakers, so Treasynx AI is launched in UK English first.
Known issues
- Splitting text on the phone. The phone's runtime splits some text into tokens differently from training: 60.9% of test prompts had at least one different split, and Thai is the most affected. In checks on a computer, the answers still passed the app's checks, but long answers were not checked this way.
- Device testing. It has run on one phone model so far. The graphics processor crashed on that phone, so the model runs on the processor instead.
- Help answers can echo hidden instructions (see Where it loses).
Training data
All training examples are synthetic: generated from the app's own code, and from examples and productivity notes that we wrote or approved. There are 100,634 training examples and 5,376 preference pairs. No user data, no text scraped from the web and no energy ratings were used, and no data from Help improve Treasynx.
More
The changelog lists each version. AI in Treasynx explains how it works in the app and what it sends, which is nothing. To report a wrong or harmful answer, tap the label under it in the app, or email support@bmvp.ai.