AI & ML
How I Built a 3.3MB Native Windows AI Desktop App with DeepSeek & Tailwind (Goodbye 150MB Electron Bloat)
Luis Hernandez DEV Community
1 views
Every time I want to build a desktop utility in 2026, the modern web ecosystem gives me the same repetitive answer: "Just bundle it with Electron!"
Don't get me wrong: Electron is an impressive engineering feat. But packaging an entire Chromium browser and Node.js runtime just to display a local database and fire off a few API calls means:
160 MB to 220 MB download installers.
500 MB to 700 MB of RAM consumption at idle.
3 to 4 seconds of cold startup lag on an average laptop.
A few weeks ago, I needed a specialized B2B cold outreach software for Windows — something that could import lead lists, craft hyper-personalized emails using modern LLMs (DeepSeek V3, Claude 3.5 Sonnet), and send them via transactional REST APIs without monthly SaaS fees.
I refused to accept a 200MB bloated binary for this.
Instead, I built LeadHunter AI Desk: a 100% native Windows desktop application with a modern Tailwind CSS dark-mode UI that weighs just 3.3 MB in a portable package and launches in 0.18 seconds.
Here is how the architecture works and how you can build similar lightweight desktop tools.
🏗️ The Anti-Bloat Architecture: Native Win32 + Edge WebView2
The secret to building ultra-lightweight desktop applications without sacrificing modern HTML5/CSS aesthetics is simple: stop distributing the browser runtime with your app.
Every modern installation of Windows 10 and Windows 11 already includes the Evergreen Microsoft Edge WebView2 runtime pre-installed and updated by Windows Update.
To orchestrate this, I used VisualNEO Win — a rapid application development (RAD) environment for Windows that compiles directly into native Win32 executables:
+---------------------------------------------------------+
| LeadHunter AI Desk |
+----------------------------+----------------------------+
| Native Win32 Core | Reactive UI Layer |
| (VisualNEO Win Runtime) | (Edge WebView2 Runtime) |
+----------------------------+----------------------------+
| * Zero Chromium bundle | * Tailwind CSS Dark Mode |
| * Local INI key storage | * Real-time KPI Dashboards |
| * Native Windows Events | * Interactive Data Tables |
| * CSV File Handlers | * Batch Selection Modals |
+----------------------------+----------------------------+
| Bidirectional Async Messaging Bridge |
| (wvPostMessage <--> OnWebMessageReceived) |
+---------------------------------------------------------+
By decoupling the UI layer from a bundled browser:
The compiled .exe is tiny.
The UI renders with full hardware acceleration via Microsoft Edge.
Memory consumption drops from 650 MB down to ~35 MB.
⚡ The Head-to-Head Benchmark
Metric
LeadHunter AI Desk
Typical Electron Outreach Tool
Download Size
3.3 MB (Portable Zip)
185 MB
Installed Footprint
~8 MB
420 MB
Startup Speed
< 0.2s (Instantaneous)
3.2 seconds
Idle RAM Footprint
~35 MB
580 MB
Pricing Model
$0 / Free & Open Source (BYOK)
$49 to $99 / month
🧠 Brains: Integrating DeepSeek V3 / R1 via OpenRouter
Instead of paying a $50/month SaaS subscription that marks up AI tokens by 500%, LeadHunter AI Desk uses a BYOK (Bring Your Own Key) model connected to OpenRouter:
You plug in your personal OpenRouter API key.
Select your preferred LLM: DeepSeek V3 / R1, Claude 3.5 Sonnet, GPT-4o Mini, or Gemini 2.0 Flash.
Use the dynamic prompt template studio with variables like {company_name}, {city}, {website}, and {my_service}.
With DeepSeek V3, generating 1,000 unique, personalized cold emails costs approximately $0.04 in direct API credits. That represents a 99.8% cost reduction compared to commercial cloud outreach platforms.
📨 Delivery: Bypassing Port Blocks with SMTP2GO REST API
If you have ever built a desktop tool that sends emails, you know the pain: residential ISPs and corporate networks routinely block outgoing traffic on Port 25, 465, and 587 to prevent botnet spam.
LeadHunter AI Desk circumvents this entirely by transmitting emails via SMTP2GO's official HTTPS REST API (Port 443):
Outgoing traffic looks like standard secure HTTPS web traffic.
Never blocked by local firewalls or anti-virus port inspectors.
Includes a configurable anti-spam delay interval (e.g., 5-10 seconds between dispatches) to protect sender domain reputation.
🛠️ Inspecting the Code & VisualNEO Win
The entire project is open source under the MIT License.
If you want to inspect how the native Win32 runtime talks to the WebView2 frontend:
; Send data to the Tailwind UI frontend
wvPostMessage "WebBrowser1" "LOAD_CONFIG|{\"model\":\"deepseek/deepseek-chat\"}"
; Handle user interactions triggered from the HTML layer
:OnWebMessageReceived
StrParse "[WebMessage]" "|" "[Action]" "[Data]"
If "[Action]" "=" "START_OUTREACH"
...
EndIf
Return
You can download the full project, edit the .pub file, and test it yourself:
📦 GitHub Repository: https://github.com/sinlios/leadhunter-ai-desk
⬇️ Pre-compiled Portable Release (3.3 MB): https://github.com/sinlios/leadhunter-ai-desk/releases/tag/v1.0.0
🖥️ VisualNEO Win IDE (Free Trial): https://visualneowin.com
💬 Final Thoughts
The web is an amazing place, and HTML/Tailwind CSS is still the most expressive UI styling system in existence. But we don't need to ship a 150MB operating system emulator just to draw a few buttons and tables on Windows.
What is your preferred approach to building desktop software in 2026? Are you sticking with Electron, experimenting with Tauri, or rediscovering native Windows tooling?
I'd love to hear your thoughts in the comments below! ⭐
Read original: https://dev.to/sinlios/how-i-built-a-33mb-native-windows-ai-desktop-app-with-deepseek-tailwind-goodbye-150mb-electron-25c7
← Previous
How I Engineered the World's First Khattak Pashto LLM (Qwen2 + LoRA)
Next →
The Web HIG: a versioned behavioral contract for humans, CI, and AI agents
Related
From Code Reviewer to Agent Manager: Lessons from 30 Days of AI-Generated Software
AI & ML
0
Dev.to (EN Zone)
Using Ctrl+Enter to Submit in OpenCode and Enter for New Lines
AI & ML
0
Dev.to (EN Zone)
7 Best Email MCP Servers and APIs for AI Agents (2026)
AI & ML
0
Dev.to (EN Zone)
Preventing Accidental OpenCode Exits with Ctrl+C
AI & ML
0
Dev.to (EN Zone)
Comments0
No comments yet — be the first