Everything. On your machine.

21+ tools — no cloud, no API key, no subscription. All in one AppImage / Installer.

⚡ Core AI 9

🧠

Local LLMQwen 3.5 · bundled

⚡

Raw ModePure inference · zero pipeline

🔒

EncrypterAES-256-GCM · text/file/dir

📄

Hybrid RAG v2BM25 + semantic · whole-doc mode

📁

WorkspaceSandboxed file ops

🎙

VoiceSTT + TTS · 5 languages

💭

Persistent memoryEpisodic + semantic

⚡

CUDA acceleration3–15× faster on NVIDIA

🎯

Prime DirectiveStanding instructions

🛠 Operations 4

👥

Build CrewLive agents · Planner · Builder · Inspector

🚀

Campaign BlasterBulk WhatsApp + email

🧰

Developer toolboxHash · minify · JSON

📅

Visual agendaCalendar + scheduling agent

📡 Services 5

💬

WhatsAppBaileys · QR · auto-reply

✈️

TelegramBot API · polling

📧

EmailSMTP · IMAP

💼

SlackBot · webhooks

🎮

DiscordBot · webhooks

🧬 Intelligence 4

🕷

Web Scraper10 URLs · batch · LMIM analysis

📇

ContactsNatural name resolution

🎨

Image generationOn-device · v2.3

🛡

Horus securityAudit trails · v3.0

⬇ Download Learn more

⬢ THE TEZCAT FOUNDATION · EVERYTHING FROM v2.1 INCLUDED

The same fire.
Sharper edge.

The full v2.1 arsenal — workspace, scraper, contacts, voice — ships inside v2.2.
Linux AppImage + Windows Installer live now. One click — everything's already running.

RAG Lite + Workspace

Drop a document. Ask anything about it. Point it at a folder and let it build.

Local embeddings · all-MiniLM-L6-v2
Smart chunking · respects headings + code blocks
MMR reranking · non-redundant retrieval
Sandboxed file ops · safe_path() blocks traversal

Just say it.

"Summarize this contract in 3 bullets." Reads the PDF, answers only from what's inside.
"Refactor main.py to use async." Reads, edits, runs — entirely within your workspace.
"What are the risks in this doc?" Suggested prompts auto-generated on upload.

Scraper + Contacts

Batch 10 URLs in parallel. Name-resolve anyone in your contacts.

Parallel fetch · 30s per URL
Basic mode or LMIM analysis mode
Natural-language contact lookup
"Send WhatsApp to Maria" — resolved automatically

Token generation · Qwen 3.5 2B tok/s

CPU only 17

GTX 1650 Ti 80

RTX 3060 120

measured · same model · same prompt

Hardened.

4-layer tool parser. Strict → embedded → relaxed → regex. No more silent failures.
Qwen3 reasoning_content fallback. Thinking models no longer return blank replies.
Cloud APIs are still optional. OpenAI, Anthropic, Groq — toggle in settings, never required.

Download v2.2 — Live Now Download manual (.md)

🐧 Linux · v2.2 Live 🪟 Windows · v2.2 Live v3 · In planning

What LMIM Does

One AI. Infinite possibilities. All tools built-in.

⚡

Raw Mode NEW

Pure LLM inference. One button strips prompts, memory, tools, and RAG. Nothing saved, ever.

✅ ~30-token identity prompt — that's it
✅ Completely ephemeral · zero history
✅ Direct llama-server call · zero pipeline
✅ Red scan line · terminal-style prompt

🔒

Encrypter NEW

AES-256-GCM for text, files, and entire directories. On-device, passphrase never stored.

✅ 100,000 PBKDF2 iterations · 256-bit keys
✅ .lmim container — documented binary format
✅ Browser + backend interoperable
✅ Batch directory mode · live progress

📄

Hybrid RAG v2 NEW

Drop a document. Ask anything. Semantic + BM25 retrieval catches concepts and exact matches alike.

✅ Whole-document mode — short docs, zero chunking
✅ BM25 + embeddings fused via RRF
✅ MMR reranking (λ = 0.65) · adaptive top_k
✅ Chunks respect headings + code blocks

📁

Workspace Sandbox

Point LMIM at a folder. Reads, writes, creates, edits. Path traversal blocked at backend.

✅ safe_path() — traversal rejected
✅ Native folder picker via Electron
✅ Live file tree in sidebar
✅ Ask LMIM about the project

🕷

Web Scraper IMPR

Batch 10 URLs in parallel. v2.2 targets article content first and filters the filler.

✅ Basic mode or LMIM analysis mode
✅ Respects robots.txt · 30s timeout
✅ Export as JSON · copy · push to chat
✅ Rate-limited · 1s between same-domain

🎙

Voice Control

Talk to LMIM. Hear it talk back. Fully on-device — no cloud, no API key, no audio leaving your machine.

✅ Whisper.cpp STT on GPU / CPU
✅ Piper TTS — 5 languages
✅ Device routing · setSinkId
✅ Auto-speak toggle

📇

Contacts

Say names, not numbers. "Send WhatsApp to Maria García" — resolved automatically.

✅ Natural-language name resolution
✅ Local address book
✅ Integrated with all messaging tools
✅ Works with Campaign Blaster

⚡

CUDA Lightning

Auto-detects your NVIDIA GPU. 3–15× faster than CPU. No config.

✅ llama.cpp + whisper.cpp on CUDA 12.x
✅ RTX 20xx · 30xx · 40xx (compute 7.5+)
✅ Falls back to CPU if no GPU
✅ CPU/GPU toggle in settings

💬

5-Channel Messaging

WhatsApp, Telegram, Email, Slack, Discord — all from one Service Switch panel.

✅ WhatsApp · Baileys · QR pairing
✅ Telegram · Bot API · polling
✅ Email · SMTP/IMAP
✅ Slack + Discord · webhooks

👥

Build Crew NEW

Type lmim build <description> in chat. Watch Planner → Builder → Inspector work in real time.

✅ Live agent orbs + colour-coded log stream
✅ Auto-generate + test + iterate · 10 attempts
✅ Inspector patches failures surgically
✅ Files land in your workspace

🐧

Cross-Platform

One file. Linux AppImage + Windows Installer live now. Bundled model. CUDA included.

✅ Linux: Ubuntu, Fedora, Debian, Arch, PopOS · Live
✅ Windows: 10, 11 — Live Now
✅ CUDA on both platforms
✅ Same features, same model, same $0

Release Log

v2.2 'RAW' is live. AppImage + Installer.

lmim-os — release log v2.2.0 · raw

Jun 13 v2.2 ⚡ TEZCAT · RAW — unfiltered power. Raw inference, hardened encryption, smarter retrieval. Live now.

Jun 13NEW⚡ Raw Mode — pure inference, zero pipeline, nothing saved. Direct llama-server call.

Jun 13NEW🔒 Encrypter — AES-256-GCM for text, files, directories. 100k PBKDF2. .lmim container.

Jun 13NEW👥 Build Crew — live agent tracking. lmim build <description> from chat.

Jun 13IMPR🧠 Hybrid RAG v2 — BM25 + semantic via RRF, whole-document mode under 16k chars.

Jun 13IMPR🕷 Scraper — targets <article>/<main>, filters filler, no more "Understood" non-answers.

Jun 13IMPRIntent routing — explicit BUILD_PREFIXES, zero false positives.

Jun 13IMPRFrontend split — dashboard separated into HTML / CSS / JS.

May 26 v2.1.1 ⬡ WINDOWS LAUNCH — v2.1.1 installer live. CPU voice bugs solved.

May 26FIX🎙 Windows Voice — solved CPU audio routing bugs. STT & TTS fully stable on Windows without GPU.

May 26IMPR💬 Chat Performance — optimized token streaming and UI rendering for smoother, faster conversations.

May 24 v2.1 ⬡ TEZCAT · SHARPENED — same fire, sharper edge. 11 new capabilities. Linux AppImage live now.

May 24NEW📄 RAG Lite — local embeddings, MMR reranking, smart chunking. Ask about your PDFs.

May 24NEW📁 Workspace — sandboxed file ops. Point LMIM at a folder, let it build.

May 24NEW🕷 Web Scraper — batch 10 URLs in parallel. LMIM analysis mode.

May 24NEW📇 Contacts — natural-language name resolution across all channels.

May 24NEW🎯 Prime Directive — standing instructions injected into every session.

May 24IMPR4-layer tool parser — strict → embedded → relaxed → regex. No silent failures.

May 24IMPRQwen3 reasoning_content fallback — thinking models no longer return blank.

May 24IMPRConsistent {ok, result, tool} schema across all 12 tools.

May 24IMPRVoice device routing via setSinkId · auto-speak for AI replies.

May 24IMPRModularized dashboard — 16 CSS/JS files replacing the 7k-line monolith.

May 24IMPROne-click Qwen 3.5 9B model download from HuggingFace.

May 24BUILD35 PyInstaller iterations → stable build pipeline. Flat layout + hook override.

Apr 24v2.0⬢ TEZCAT — major release on Linux. Bundled model + CUDA + toolbox.

Apr 24NEWCUDA acceleration — llama.cpp + whisper.cpp on GPU. 3–15× faster.

Apr 24NEWDeveloper toolbox — SHA hash, CSS/JS minify, JSON validator, all in-app.

Apr 24NEWCalendar agent — "schedule X tue 3pm" books + WhatsApps confirmation.

Apr 24NEWService Switch — start/stop all 5 messaging daemons from GUI.

Apr 24NEWCampaign Blaster — bulk WhatsApp + email campaigns from CSV lists.

Apr 24IMPRBundled Qwen 3.5 in AppImage — open it and it works.

Apr 24IMPRWhisper STT on GPU — ~0.3s transcription on RTX 3060.

Apr 24IMPREN/ES/PT/FR/DE TTS via Piper — five languages bundled.

↓ Download v2.2 — Go RAW

What's Next?

v2.2 is live on both platforms. v2.3 and v3 are already taking shape.

⬢

Launching now

v2.2 · Tezcat · RAW

Unfiltered power. Raw inference, hardened encryption, smarter retrieval, live agents. All tried, all working.

⚡

Raw Mode Pure inference, zero pipeline overhead. Ephemeral by design — nothing saved.

🔒

Encrypter AES-256-GCM for text, files, directories. Browser + backend interoperable.

👥

Build Crew Planner → Builder → Inspector with live agent tracking. Triggered from chat.

🐧 + 🪟 · Live Now

⬢

v2.1 · Tezcat · Sharpened

The foundation. RAG, workspace sandbox, web scraper, contacts, voice on GPU — all carried forward into v2.2.

✅ Shipped Linux + Windows · May 2026
✅ 1,000+ downloads, 20+ countries
✅ Every feature still included

✅ Shipped

🎨

v2.3 · Image Gen

On-device Stable Diffusion via sd.cpp with CUDA. Describe it, see it. No API key. The stub tab goes live.

✅ sd.cpp + CUDA acceleration
✅ Chat-native generation
✅ Quantized SD-Turbo model

🔄 Target: Q3 2026

🛡

v3.0 · Horus Engine

Enterprise-grade security. Code scanning, immutable audit trails, policy enforcement, multi-user RBAC.

✅ Static analysis on Builder output
✅ Tamper-evident action ledger
✅ Role-based tool access

🔄 Target: Q3 2026

🍎

macOS Support

Apple Silicon + Intel. Same one-file philosophy. No Homebrew, no Xcode required. Metal acceleration on M-series chips.

✅ Apple Silicon (M1/M2/M3/M4)
✅ Intel Mac fallback
✅ Metal GPU acceleration

🔄 Planning

🔌

Plugin System

Third-party tools without core changes. Drop a manifest, register your tool, LMIM discovers it automatically.

✅ Manifest-based registration
✅ Sandboxed execution
✅ Community marketplace

🔄 v3.0

Ready to go RAW?

One file. Everything included — bundled model, CUDA, no cloud. Linux live now · Windows June 12.

1,000+Downloads

20+Countries

19+Tools built-in

🐧 Linux AppImage — v2.2 'Tezcat · RAW' x86_64 · v2.2.0 · Model + CUDA bundled · GPG signed 🪟 Windows Installer — v2.2 'Tezcat · RAW' .exe · v2.2.0 · Model + CUDA bundled · AES-256 Encrypter

🐧 Linux: chmod +x LMIM_OS_2.2.0_Tezcat_RAW.AppImage && ./LMIM_OS_2.2.0_Tezcat_RAW.AppImage
🪟 Windows: Download LMIM_OS-Setup-2.2.0.exe, double-click, follow the wizard.

Install via terminal:

🐧 Linux:

wget https://lmim.tech/downloads/LMIM_OS_2.2.0_Tezcat_RAW.AppImage && chmod +x LMIM_OS_2.2.0_Tezcat_RAW.AppImage && ./LMIM_OS_2.2.0_Tezcat_RAW.AppImage

🪟 Windows: curl -L -o LMIM_OS-Setup-2.2.0.exe https://lmim.tech/downloads/LMIM_OS-Setup-2.2.0.exe && start LMIM_OS-Setup-2.2.0.exe

Download manual (README.md)

Other options:

v2.2 console · Features · Changelog · Contact

🐧 Linux AppImage (stable) + 🪟 Windows Installer (stable) · Model + CUDA bundled on both · 📄 AGPL-3.0 License · No telemetry · GPG signed

Built with LMIM

Local AI in the Real World

Not demos. Not prototypes. Production deployments running LMIM's stack where privacy and offline reliability aren't optional.

🦊

Happy Fox App

Education · Mexico

A full English learning platform for kids — AI pronunciation scoring, on-device tutoring, and 155 weeks of curriculum. Runs entirely on the teacher's machine over LAN. Zero cloud.

🎙 Whisper STT — real-time pronunciation scoring

🔊 Piper TTS — instant audio for every word

🤖 Qwen LLM — Foxy AI tutor, on-device

🔒 100% offline — LAN only, no data leaves the school

#EdTech #LocalAI #Privacy #Offline

stack: llama.cpp · whisper.cpp · piper · sqlite · electron

Build yours with Enterprise →

Building something with LMIM? Tell us about it →

Everything. On your machine.

LMIM OS v2.2 — Unfiltered power.

Local AI That
Actually Does Things

Unfiltered power.
Run it yourself.