MindMopAI agent (4-bit)
The small on-device model inside MindMopAI, a free brain-dump app for
Windows and Mac. It does two jobs, and runs entirely on your computer through llama.cpp. Nothing
you say is sent anywhere.
- Sorting. You say whatever is on your mind, and it splits it into separate items and sorts
each one: task, bill, exam, idea, note, shopping, read, social or appointment.
- Mai, the in-app helper. It answers questions about MindMopAI from a help sheet the app
provides, reads your own items and settings to answer questions about them, offers actions
(for example a theme picker, or "mark this done" for you to confirm), and reviews your task list
with practical suggestions.
Files
| File |
Version |
SHA-256 |
mindmop-agent-r4-q4.gguf |
Round 4: sorting and Mai (current) |
1926c2eba1e99e94c7a60ebda7d70841e0d8375f2d6c070ff8bf847685b1ce29 |
mindmop-agent-q4.gguf |
Round 2: sorting only (kept for older app versions) |
e1d89e85f63d616fc81d8b2a67b785e65c0de91018d4ab712d9401c1c039bb9d |
Details (round 4)
- Base model: Qwen2.5-1.5B-Instruct
- Fine-tuning: one LoRA trained on both jobs together (about 2,000 sorting examples and about
2,000 helper examples, all synthetic), then merged
- Quantization: Q4_K_M GGUF (about 1 GB)
- Sorting accuracy: 87% of 200 held-out test thoughts sorted exactly right (every item, kind
and date), with dates added by MindMopAI's rule-based parser
- Helper accuracy: 39 of 43 held-out test questions answered correctly (app questions, the
user's own data, actions and task reviews)
- Speed: about 1.3 seconds per thought on a laptop processor, no graphics card needed
Output
For sorting, the app constrains every answer with a grammar, so the model can only reply in this shape:
{"items": [{"text": "tomorrow pay the rent", "kind": "bill"}, {"text": "tomorrow buy groceries", "kind": "shopping"}]}
As the helper, it writes a short answer and may end it with an action tag such as [[do:theme]];
the app turns tags into buttons, and nothing changes until the person presses one. MindMopAI checks
every answer before using it and falls back to its rules if a check fails.
Licence
Apache 2.0, the same as the base model.
