OpenJev Runs Decision Models Locally in Your Browser

Developers and researchers now have access to OpenJev, a browser-based framework that runs local AI decision models entirely on client-side hardware. The initiative eliminates backend dependencies by loading model weights directly from Hugging Face and caching them within the browser, ensuring that all inputs and computations remain strictly local. The platform is designed to evaluate how decision models process information by comparing two distinct inference pathways. The first method extracts normalized option probabilities directly from the model logs, while the second requires the model to generate those same distributions sequentially as JSON text. Both approaches execute on the same loaded instance to prevent GPU contention, allowing users to measure wall-time ratios and token generation speeds accurately. The tool supports multiple model sizes to accommodate varying hardware capabilities. MiniCPM5 2B is selected by default, with Qwen3 0.6B recommended for mobile or lower-end devices, and Qwen3.5 4B offered for higher accuracy requirements. Benchmark results indicate a clear performance trajectory, with the 4B model achieving approximately 81 to 84 percent across safety, native BF16, and type-safe evaluation subsets. These figures closely align with published Jev baselines, demonstrating that quantized browser builds can maintain robust decision-making capabilities. First-time users should anticipate initial loading times of several minutes, contingent upon their GPU capacity and network connection. By providing a transparent, privacy-first environment for testing local AI decision systems, OpenJev bridges the gap between theoretical model architecture and practical client-side deployment. The framework enables engineers to benchmark inference latency, validate model reliability, and optimize deployment strategies without transmitting sensitive data to external servers.