openjev

SemIf is an independent research project formerly called OpenJev. It explores semantic if-statements from open models directly in the browser and explicitly states that it is not affiliated with or endorsed by TypeSafe. The project compares two local decision paths: reading option probabilities directly from model logits and asking the model to write the same distribution token by token.

Key Features

• Run the experiment in a browser with no backend; inputs stay on the page and model weights remain in the browser cache.

• Choose among local open-model sizes, including Qwen3 0.6B, MiniCPM5 2B, and Qwen3.5 4B, with MiniCPM5 2B selected for desktop by default.

• Load a model through WebGPU and measure download, model-load, warmup, direct-readout, and generation timing on the user’s own GPU.

• Define a state, a decision question, and 2–20 allowed options, then run both the direct-logit and token-generation paths on the same model.

• Inspect choice probabilities, JSON probabilities, first-token timing, total timing, input/output timing, and the wall-time ratio.

• Use the measured model-quality table and implementation notes to understand checkpoint size, accuracy comparisons, quantization, and the wllama-based browser runtime.

How It Works

Select a model and load it once. Enter the application state, question, and allowed options, then run both methods. SemIf executes the direct readout first and generation second on the same loaded model so the paths do not contend for one GPU. The results panel exposes the two probability outputs and measured timings.

Use Cases

SemIf is suited to developers and researchers studying local structured decisions, semantic if-statements, probability readouts, autoregressive JSON generation, WebGPU inference, quantized open models, and real local latency. It is a browser experiment rather than a hosted production API.

Pricing and Availability

The site presents the project as free to try with no waitlist. The model download happens only after the user clicks the load button; the first load can be several gigabytes and may take minutes depending on the selected model, network, and GPU.

Screenshots