A LITTLE MORE CLARITY

Set up Minth for Windows

New name: Minth. Developed by James van Oosterwijk. Official website: minthai.pages.dev. Install version 1.0.88 on each PC; older installations do not update automatically. Existing chats, settings, encrypted credentials, voices and models keep their legacy storage paths; no data migration or new download is needed for the rename. The green symbol and white interface remain. Explicit YouTube commands now accept “open christmas hits on youtbe”, “you tube” and “you-tube”; this is bounded command matching, not improved speech recognition.

Windows 10/11 x64. 8 GB RAM recommended; at least 2 GB free disk for local AI, plus app/voice storage. No account or API key.

1. Download and verify

Get Minth-Setup-1.0.88-x64.exe from this site's release link. Compare Get-FileHash .\Minth-Setup-1.0.88-x64.exe -Algorithm SHA256 with the release checksum. This app is unsigned. Windows may warn; verify source and checksum. Never disable antivirus, SmartScreen or other security protections.

2. Install Minth

Open the setup EXE. The per-user installer creates Minth shortcuts without administrator rights, does not automatically launch Minth, and does not download models. Close Minth before updating. Existing chats, settings and keys remain in the separate Mint Chat profile, not the app folder; inactive cloud credentials are preserved but unused.

3. Set up local AI

Open Your workspace → Connection → Local AI. Qwen2.5 1.5B Instruct Q4_K_M (Apache 2.0) runs on CPU through llama.cpp b11655. Approve about 1.14 GB from Hugging Face/GitHub if it is not already installed. Existing verified Qwen2.5 files are reused without another download; Qwen3.5 files stay on disk but are not loaded. Integrity checks remain required. No automatic downloads, account, API key, AI usage bill or cloud fallback. Download hosts receive normal requests/IP, not chats.

This release restores the smaller model after slow responses with 4B. Install Minth 1.0.88 on each PC; installed older versions do not update automatically. Once set up, text inference can work offline. The model preloads when Minth opens, using RAM. One rollback test greeting took about 1.5 seconds after a 14-second preload. These are samples, not a speed guarantee for other laptops. Dutch and factual quality are limited; best in English, not ChatGPT-level. Text-only, no image understanding. Gemini, ChatGPT/OpenAI and ElevenLabs are disabled in this release.

4. Supported local actions

Tools & voice → Desktop actions: Ask first is the new-profile default; optional Automatic runs supported requests immediately. Up to 10 validated app/site launches per request, simple house/rocket/smiley/tree/flower/sun drawings and Windows output volume 0–100%. “Put volume to 11” means 11%, verified by Windows read-back. Microphone, mute and per-app levels stay unchanged. Drawings create uniquely named PNGs in Documents/Mint Chat Creations and request Paint; existing files are never overwritten.

Explicit Google search commands send only that query to Google. A first-result lookup uses an isolated temporary Chrome profile, never logged-in tabs. Verification is not bypassed: blocked lookup opens results and says no result was clicked. “Play christmas hits on YouTube” checks a public video ID and opens a watch page with autoplay requested; consent/sign-in/Play may still be needed. Playback is never confirmed. Lookup failure opens only search results. Opened sites make normal network requests; AI does not read them.

No arbitrary commands, file deletion, unrestricted reading, typing/clicking, app closing or media downloads. Optional “Open Minth Dev” only launches an already installed fixed personal Minth path; it does not install or copy anything.

YouTube and volume together

“Play christmas hits on youtube and put my volume to 100” now supports both actions in one request. Both complete clauses must validate first, in either order (English/Dutch). Ask first previews each; saved Automatic runs both without the model. 100% can be loud. YouTube may need consent/sign-in/Play; playback is not guaranteed. Unsupported, extra, negative or attached clauses do not partially execute. Only this pair is supported here, not arbitrary chaining.

Optional automatic web excerpts

Tools & voice → Allow local AI web excerpts is off by default. Enable it and Save preferences: current short public-topic questions are then searched automatically, without approval each time. Saved choices are retained on update. DuckDuckGo receives only the current question (plus generic cast/actor/official terms for casting questions), not chat history, attachments or custom instructions. Private/credential-like and ambiguous follow-ups are refused. Uncheck and Save to disable. This is separate from the desktop Automatic/Ask first policy.

Minth reads a bounded public search page and can select an exact quote locally. Results have source links and retrieval time, but are not full-page verification or guaranteed current facts. No linked-page fetching, cookies, scripts, retries, paid API or cloud AI. Failed selection shows labeled excerpts; failed/blocked lookup reports that no current answer was verified. Web-excerpt UI and local quote selection were retested with Qwen2.5. Availability and speed are not guaranteed.

5. Local voice

Voice settings → Set up free voice. Approve separate component downloads (several hundred MB) and keep Minth open. No scripts to run manually, no voice API key or credits. Local recognition and Piper English/female Flemish Dutch speech are available. No cloud/system/browser speech fallback. Dii is not bundled. Turn-based voice, not realtime; headphones recommended. Filtering does not identify only your voice. Components retain upstream licenses.

Backups, updates and uninstalling

Your data → Back up workspace exports private chats/attachments/notes, not API keys. Keep backups private. Close Minth before running a newer installer. Uninstall preserves the separate profile. The optional --mint-public-test flag uses a separate Mint Chat Public Test profile/port; normal installation uses Mint Chat. No automatic updater.

Known limits

Experimental unsigned software. No security or accuracy guarantee, compatibility guarantee or support SLA. Check important output. Windows x64 only; no mobile/Mac build. Older versions had different cloud/privacy behavior; these instructions apply to 1.0.88.