ThinkWell v0.2.3 — Benchmarks, color themes, rich replies & export

ThinkWell is a private, offline AI companion that runs local GGUF models entirely on your device — no cloud, no account, nothing leaves your phone.

This is a large rollup of everything since v0.0.20. APK attached below (Android, arm64 + x86_64).

⚠️ Prerelease (v0.2.x). On-device AI can warm the device, use battery, and run slower or less accurately than the cloud — it depends on your hardware and the model you pick.

✨ Highlights

🧠 Models & downloads

• Benchmark-based "Powerful" sort — models ranked by the average of their published MMLU, IFEval, and GSM8K scores, with an ⓘ info popup on every model showing each score and a tappable source (or noting when a rating is estimated).
• Color-coded RAM & download-size tags (green = light → red = heavy) and a device-fit warning that asks for confirmation before downloading a model heavier than your phone's RAM.
• Filter the catalog by input type — All · Text · Image · Audio.
• 6 new models — Qwen3 1.7B & 4B (2507), SmolLM2 1.7B, IBM Granite 3.1 2B, and GLM-Edge 1.5B & 4B.
• Downloads now run one at a time, so a model never shows "Use" before it's actually finished.

💬 Chat & replies

• Rich replies, fully offline — real LaTeX math (KaTeX), syntax-highlighted code (highlight.js), and Markdown tables, rendered reliably across models.
• Each reply is tagged with the model that produced it, and shows an animated generating indicator while it works.
• Copy, Share, and Export any reply — export as plain text, Markdown, Word (.doc), or PDF; files are named with the date/time and carry a small "ThinkWell" line. PDF keeps full formatting with page margins.
• Simple/Advanced interface mode hides technical settings for newcomers.

🎨 Appearance

• Light / Dark / System themes, plus five neutral color palettes — Warm Beige (default), Sage, Sky, Lavender, Rose — for both light and dark.

🌍 Languages

• Fully translated into 30 languages, including right-to-left layout for Arabic, Farsi, and Hebrew — model descriptions, benchmark notes, sort labels, and workflows are all localized.

🧭 Onboarding & data

• A first-run + replayable tutorial (updated to cover the new features), themed in-app dialogs, Backup & Restore, and per-chat / clear-all history controls.

🗒️ Full changelog (since v0.0.20)

Expand

v0.2.3 — Three reply actions (Copy / Share / Export), each adding a "ThinkWell" line; date-time export filenames; file-only export sheet; new tutorial slide; README update.
v0.2.2 — Export menu Text/File sections + ThinkWell footer; serialize downloads (fix premature "Use"); PDF page margins; Word math as readable text; render fenced Markdown tables; softer math fallback.
v0.2.1 — Translate the remaining English-only text (model descriptions, benchmark notes, sort labels, workflow names) into all 30 languages.
v0.2.0 — Benchmark-based Powerful sort; per-model ⓘ capability popup with sourced scores; color-coded RAM/storage tags; device-RAM download warning; five neutral color palettes; calmer default theme.
v0.1.0 — 6 more models (Qwen3 1.7B/4B, SmolLM2 1.7B, Granite 3.1 2B, GLM-Edge 1.5B/4B); input-type filter; reply model tag; copy/export (text/Markdown/Word/PDF); animated generating indicator; cross-model math rendering.
v0.0.28 — Offline KaTeX math, highlight.js code, and Markdown tables in replies.
v0.0.27 — Simple/Advanced mode; no-model reply nudge; cleaner inline math; RTL tutorial paging fix; developer link.
v0.0.26 — Full 30-language translations; themed in-app dialogs; logo on tutorial slide 1; RTL single-restart fix.
v0.0.25 — 30-language UI; right-to-left support for Arabic, Farsi, Hebrew.
v0.0.24 — First-run + replayable app tutorial.
v0.0.23 — Light & Dark themes (System/Light/Dark).
v0.0.22 — Backup & Restore conversations.
v0.0.21 — Clear-all + per-chat erase; color-coded modality tags; redesigned About section.

---

🧠 Think well, on your own terms.

v0.0.20

ThinkWell is a private, offline AI companion that runs local GGUF models
entirely on your device via llama.rn — no account, no server, no data leaves
your phone.

✨ New in v0.0.20

• First-run welcome notice — a one-time heads-up that on-device AI can warm
  the device, use battery, and run slowly (normal, not a fault of the app or your
  device), that speed and quality depend on your hardware and chosen model, and
  that offline models are smaller than cloud AI so they can be inaccurate.

🔊 Also recently added

Audio input (v0.0.19)

• Attach a WAV/MP3 clip or record from the microphone and ask a
  speech-capable model about it — fully on-device.
• New Voxtral Mini 3B (Audio) model: transcribes and answers questions
  about your audio.
• Mic recordings are captured as 16 kHz mono WAV (the format the engine accepts).

Models & rendering (v0.0.18)

• Re-added Mistral 7B Instruct v0.3 and Phi-3.5 Mini — now coherent on
  the current engine.
• Fixed Markdown lists rendering one word per line in long replies.

📥 Install

• Android only. Download ThinkWell.v0.0.20.apk below and install (you may need
  to allow installs from unknown sources).
• Models are not bundled — pick one in the Models tab and it downloads on
  first use.

⚠️ Notes

• Early prerelease (v0.0.x) — expect rough edges.
• On-device models are far smaller than cloud AI; double-check anything important.
• Voxtral Mini sometimes adds a stray "I can't process audio" line even while
  answering correctly — a model quirk, not an app bug.