Local Qwen3.8 Model Builds and Debugs 3D Browser Game via Agent Harness
A developer demonstrated using a locally hosted quantized model (Qwen3.8-Flash-Next via Intel AutoRound W4A16) within the OMP agent harness to create and self-debug a 3D browser game from a short, vague prompt. Over a three-hour run, the agent autonomously controlled two browser instances simultaneously to visually and functionally test, debug, and refine HTML and JavaScript code. This demonstration illustrates the growing feasibility of running complex, multi-hour autonomous software engineering workflows entirely on local hardware using open-weight quantized LLMs. It shows how agent harnesses equipped with browser execution capabilities allow smaller models to fix their own mistakes through iterative feedback loops. The setup ran locally across four NVIDIA V620 GPUs, achieving generation speeds of approximately 2,000 prefill tokens per second and 70 decode tokens per second using Intel AutoRound W4A16 quantization. The OMP agent harness coordinated state tracking and active browser execution to continuously catch runtime and rendering errors over three hours.
## BACKGROUND
Quantization methods such as Intel AutoRound compress LLM parameters to low bit-widths (e.g., 4-bit weights) to reduce memory usage while preserving model capabilities. Agent harnesses like OMP (Oh My Pi) wrap language models with external tool execution, state management, and browser automation, allowing the AI to run code and observe outputs in closed-loop feedback environments.