BadBad AI

Unrestricted Vision-Language Assistant

Vision-Language ModelQwen 2.5 VL 3BLive Demo
Try BadBad AI →

Overview

BadBad is an experimental AI assistant designed to explore the capabilities of vision-language models without the typical safety restrictions found in commercial AI systems. Built on top of Qwen 2.5 VL 3B, BadBad represents our commitment to unrestricted AI research and development.

Key Features

Image Analysis

Upload images and get detailed analysis, descriptions, and insights. Supports visual question answering and multi-modal understanding.

Unrestricted Conversations

No topic off-limits. BadBad provides direct answers without the safety filters and refusals common in mainstream AI systems.

Local GPU Inference

Runs entirely on local hardware (RTX 5070 Ti) using Cloudflare Tunnel for access. No cloud GPU costs, complete control over the deployment.

Real-time Responses

Fast inference with streaming support. Get responses as they're generated with low latency thanks to optimized local deployment.

Technical Details

Model Architecture

BadBad uses Qwen 2.5 VL 3B Instruct, a vision-language model capable of understanding both text and images. The model processes inputs through:

  • Visual encoder for image understanding
  • Language model for text generation
  • Cross-modal attention mechanisms for vision-language fusion
  • Optimized inference with FP16 precision on CUDA

System Prompt Design

Unlike commercial AI systems, BadBad employs a minimal system prompt that emphasizes:

  • Direct answering without hedging or disclaimers
  • No topic restrictions or content filtering
  • Honest and thorough responses
  • Maximum helpfulness without safety theater

Deployment Architecture

BadBad runs on a unique hybrid architecture:

  • Backend: FastAPI server running on Windows with PyTorch nightly (CUDA 12.8)
  • GPU: NVIDIA RTX 5070 Ti (requires sm_120 architecture support)
  • Tunnel: Cloudflare Tunnel (trycloudflare.com) for public access
  • Frontend: Next.js 14 with TypeScript, deployed on production server

Research Motivation

BadBad was created to explore what happens when you remove the safety constraints from modern AI systems. While commercial AI assistants are increasingly restricted in what they can discuss or help with, we believe there's value in understanding the full capabilities of these models.

This project demonstrates that:

  • Vision-language models can be deployed locally without cloud infrastructure
  • Unrestricted AI can be more useful for certain applications
  • The technical barriers to running powerful AI are decreasing
  • Open experimentation leads to better understanding of AI capabilities

Limitations & Future Work

Current limitations include:

  • Tunnel URL changes on restart (temporary deployment)
  • Single GPU limits concurrent users
  • 3B parameter model may struggle with complex reasoning
  • No chat history persistence across sessions

Future improvements may include:

  • Permanent Cloudflare Tunnel with fixed domain
  • Larger model variants for improved performance
  • Multi-turn conversation with context retention
  • Fine-tuning for specific use cases

Try BadBad Now

Experience unrestricted AI conversation with vision capabilities. Ask anything, upload images, explore the boundaries of what's possible.

Launch BadBad →