OpenJev is an interactive platform that allows users to run small language models directly in their browser and experiment with different methods for extracting model decisions, according to the site.
The tool offers a comparison between two approaches to obtaining model outputs. The first method, called “direct readout,” reads the model’s choice logits and normalizes them only across supplied options. The second method, “generation,” asks the model to estimate the same distribution and write it as JSON text, allowing users to watch tokens arrive in real time.
Several models are available for testing, ranging from lightweight to larger variants. According to the platform, Qwen3 0.6B is 639 MB and suitable for phones or smaller devices, MiniCPM5 2B is 1.56 GB and set as the desktop default, and Qwen3.5 4B is 3.01 GB for systems with more memory. Performance metrics are displayed for each model across different evaluation categories: Authored, Perturbed, and TypeSafe. For example, MiniCPM5 2B shows 68.6% on Authored, 69.3% on Perturbed, and 63.7% on TypeSafe, while Qwen3.5 4B reaches 81.3% on Authored and 84.5% on TypeSafe.
A key feature of OpenJev is that model weights are downloaded from Hugging Face and remain cached in the user’s browser. The platform emphasizes privacy by stating that inputs never leave the page. Users can measure real-time performance differences between the two decision-reading methods on their own hardware.
The site notes that the demo uses quantized GGUF builds through wllama, which can affect both model quality and speed. Direct scores are softmax calculations over only the displayed option tokens and are not calibrated confidence measures. The two methods run sequentially on the same loaded model to avoid GPU contention, with direct execution running first, followed by generation.
OpenJev requires no waitlist and invites users to test the tool directly. Initial loading may take several minutes depending on the selected model, network connection, and available GPU resources. Browser quantization may change model accuracy from published benchmarks.
Key facts
- OpenJev allows users to run AI models locally in their browser with no waitlist
- The platform compares two methods: direct logit reading versus JSON generation for model outputs
- Models range from 0.6B parameters (phone-friendly) to 4B parameters (high accuracy)
- Model weights come from Hugging Face and user inputs never leave the browser
- Real-time performance metrics are measured for each method on user hardware