DEV Community

Cover image for Training Nova-Mini V1: Building a 57M Parameter Edge AI Model for Desktop Automation
Achyut Srivastava
Achyut Srivastava

Posted on

Training Nova-Mini V1: Building a 57M Parameter Edge AI Model for Desktop Automation

What We Built Today

We successfully trained LuxurAI Nova-Mini V1 — a 57,192,128-parameter edge AI model specialized for desktop automation and screen understanding. Unlike wrapper-based approaches, Nova-Mini is trained from first principles with Chain-of-Thought reasoning baked directly into its architecture.

The Problem We Solved

Existing desktop automation tools rely on brittle pixel matching or API wrappers. Nova-Mini V1 introduces native visual-language understanding, allowing it to:

  • Interpret screen layouts as structured hierarchies
  • Ground OCR text in visual context
  • Plan multi-step workflows with tool delegation

Architecture Highlights

  • Dual-Tower Encoder: Unifies screen patches (via Neo v0.1 vision cluster) and text tokens into a joint embedding space
  • Tool Calling: Delegates execution to sandboxed Python/Shell REPLs rather than hallucinating code
  • Training Rig: Dual Tesla T4 GPUs with gradient checkpointing (batch size 8, seq len 2048)

Current Benchmarks

  • Math (Python REPL): 9.5/10
  • Shell Scripting: 8.5/10
  • Algorithm Coding: 7.5/10

How We Can Make This Better

We're actively working on:

  • High-DPI multi-monitor perception
  • Lower-latency Windows UIA integration

Would love feedback from the community! Try the beta at LuxurAI Cowork.

Top comments (0)