Curious 2B
Curious 2B is the on-device model used by CuriousLM. It is Qwen3.5 2B fine-tuned on the tasks CuriousLM's assistant performs on Android: answering questions from the user's calendar and messages, adding, moving and removing calendar events and setting reminders through tool calls, reading incoming messages and notifications (what a message asks, whether a reply is needed, bills, bookings, deliveries and scams), summing up notifications, and writing the assistant's morning and evening notes and reply drafts.
The model runs locally with llama.cpp. CuriousLM supplies the calendar and notification data and runs the tools; nothing is sent to a server.
Releases
| Release | File | Size | SHA-256 |
|---|---|---|---|
curious-2b-261012 (12 October 2026, current) | curious-2b-261012-Q4_0.gguf | 1.21 GB | 9ad72540cc2b400fd751bf56200f8ac33d49f5df45ea5e20131538bc8579110c |
curious-2b-261011 (11 October 2026) | curious-2b-261011-Q4_0.gguf | 1.21 GB | ed940680c4ed258ee14e0fa849ed12e51b559953eb432f99dec377e65545a9e7 |
curious-2b-261010 (10 October 2026) | curious-2b-261010-Q4_0.gguf | 1.21 GB | e106382634f3ec5319fb7fa5a6d0bd28e8c5fb1d031d8ae29b8e8866fc743f0f |
curious-2b-261009 (9 October 2026) | curious-2b-261009-Q4_0.gguf | 1.21 GB | 42b4c223434842288c999a447472df33d118169b60eeda55622d0c01866ca333 |
curious-2b-2610 (8 October 2026) | curious-2b-2610-Q4_0.gguf | 1.21 GB | 45840b88cc63279df6b42c857832068d85c6f9a13745cebcb64fc792af45320d |
Evaluation
Held-out cases, both releases under the same prompts and settings (CuriousLM's calendar tools: add, move, remove and remind).
Phone and calendar
| 261011 | 261012 | |
|---|---|---|
| Calendar edits: remove or move an event (13) | 30.8% | 69.2% |
| Reminders in English, Spanish, French, German and Portuguese (24) | 95.8% | 100.0% |
| Plain instructions with tools offered and none needed (36) | 83.3% | 100.0% |
| Turns passed (655) | 88.4% | 86.1% |
| Requests with times said in words, marked and cleared days, day of the month, people in titles (580) | 85.3% | 82.1% |
| Questions about a past day (150) | 86.0% | 80.7% |
| Reminder and alarm times in English, Spanish, French and German, from MTOP (1,050) | 94.0% | 86.5% |
| Says it cannot read the phone | 0.0% | 0.3% |
| Claims an action that was not taken | 0.3% | 1.5% |
Notifications
| 261011 | 261012 | |
|---|---|---|
| Notification summaries for the away sheet and home (275) | 76.7% | 92.7% |
| The week ahead in one or two sentences (42) | 52.4% | 59.5% |
| Reading: standard set (519 tasks) | 97.5% | 98.7% |
| Reading: hard set, receipts, deliveries, friends' plans, scams (599) | 95.8% | 97.2% |
| Reading: real text messages (300) | 78.7% | 79.0% |
| Right action for a notification (210) | 84.8% | 83.3% |
The assistant's writing
| 261011 | 261012 | |
|---|---|---|
| Morning and evening notes, what came in, reply drafts (250) | 79.2% | 74.8% |
| The same, in the app's current writing style (250) | 83.2% | 74.0% |
| Home briefing leads with what needs the user (120) | 92.5% | 91.7% |
| Home briefing under the app's current lead rule (120) | 94.2% | 91.7% |
General chat
| 261011 | 261012 | |
|---|---|---|
| Knowledge questions (360) | 98.9% | 96.9% |
| Identifies itself correctly (72) | 100.0% | 100.0% |
| Open answers: judged false facts per answer (1,009) | 2.8 | 2.9 |
Use
In CuriousLM: Models, then Curious 2B.
With llama.cpp:
llama-server -m curious-2b-261012-Q4_0.gguf --jinja
The tool definitions and prompts the model was tuned on are CuriousLM's own. For general function calling outside the app, Qwen3.5 2B scores higher.
Details
| Base model | Qwen3.5 2B |
| Method | LoRA fine-tune, merged |
| Format | GGUF, Q4_0 |
| Languages | English, Spanish, French, Portuguese, German, Italian, Dutch, Chinese |
Licence
Apache 2.0. Fine-tuned from Qwen3.5-2B, Copyright 2026 Alibaba Cloud, Apache 2.0. Training data includes MASSIVE (Amazon) and When2Call (NVIDIA), both CC BY 4.0; the SMS Phishing Dataset (Mishra and Soni, Mendeley Data, doi:10.17632/f45bkkt8pr.1) and the SMS Spam Collection (Almeida and Hidalgo, UCI), both CC BY 4.0; and spoken date and time phrases from Recognizers-Text (Microsoft, MIT) and Duckling (Meta, BSD 3-Clause).