ThinkWell
A private AI companion that runs entirely on your device.
A complete AI assistant that never leaves your device.
Fully Offline AI Chat
Real-time, token-by-token conversations powered by a language model running locally on your phone.
Private by Design
No cloud, no accounts, no analytics. Your conversations and settings stay on your device, except a reply you choose to report.
Rich, Readable Replies
Renders Markdown with LaTeX math, syntax-highlighted code, and tables — all bundled to work offline.
On-Demand Model Catalog
38 curated, phone-friendly models across thirteen families (Gemma, Llama, Qwen, DeepSeek, Phi, Mistral, SmolLM, GLM-Edge, Granite, Hermes, LFM, MiniCPM and Voxtral), each tagged with its size and RAM needs.
Vision & Audio Input
Attach a photo or record speech and ask the model about it — multimodal, and still fully on-device.
Compose Music
Describe a mood, genre or instruments and MIDI-LLM writes a piece on your phone. Play it in the chat or share it as a MIDI file.
Read Aloud
Hear any reply, or have every reply read automatically, in your phone’s own offline voices. Nothing is sent online to be spoken.
Bring Your Own Model
Import any local GGUF model file and run it alongside the built-in catalog.
Export Anywhere
Save any reply as plain text, Markdown, Word, or PDF — with math, code, and tables preserved.
Guided Workflows
Twelve structured templates — Decide, Plan, Debug, Explain, Translate and more — each configurable before you run it: target language, tone, length, audience or review focus.
Build Your Own Workflow
Duplicate one, or build a new one from scratch with a step-by-step wizard. Share it as a file and import someone else’s. A workflow can ask for a picture or an audio clip as well as text.
Send to Another AI
Hand a whole conversation to another assistant as a compact briefing. Long chats are condensed by a model already on your phone, offline — no cloud service writes the summary.
Per-Model Guidance
ThinkWell warns you, in your own language, when the chosen model is documented to handle a workflow badly, and each model’s info sheet lists the languages its maker officially supports.
Memory-Safe Runs
ThinkWell watches free memory while a model runs and stops it before the system would close the app, warns you before starting a model that will not fit right now, and tags models that ran out of memory on your device.
Report a Reply
Flag a harmful or wrong reply from right under it. The report goes to the SeMo Lab team for review, and nothing else leaves your phone.
Guided App Tour
A short spotlight tour walks you through every main feature on the real screens, and a What’s New card after each feature update tells you what changed.
Pin & Rename Chats
Keep the conversations you return to at the top of the list, under names that mean something to you.
Long-Term Memory & Backup
Optional on-device memory recalls facts across chats; back up and restore your chats as a single dated .thinkwell file.
30 Languages, Including RTL
A fully translated interface in 30 languages, with right-to-left support for Arabic, Farsi, and Hebrew. A one-time note explains that the setting changes the app, not the language the AI replies in.
Themes & Simple/Advanced Modes
Light or dark with five colour palettes and an AMOLED black option for OLED screens, plus Simple and Advanced modes for every kind of user.
Benchmark Your Device
Run a short on-device speed test to see your device tier and real tokens per second — afterwards every model card shows the speed you can expect from it.
Downloads That Finish
Model downloads carry on if the app is killed and resume where they stopped, falling back to a mirror when the source is slow or unreachable.
See ThinkWell in action.
