OpenJev brings open-weight decision models to browser
OpenJev is an in-browser demo and open-source project created by Theo Lee that replicates the core functionality of TypeSafe's proprietary Jev decision model using open-weight models such as MiniCPM and Qwen. Running entirely on client hardware via WebGPU and wllama without backend servers, OpenJev allows users to compare two distinct decision-making approaches on identical prompts: reading normalized choice logits directly in a single forward pass versus autoregressively generating structured JSON token by token.
Generating conversational JSON strings just to make a discrete classification decision is pure architectural overhead—direct logit extraction represents the natural future of agent routing and guardrails, even if probability calibration on edge devices requires careful handling.
- –Single-pass logit extraction eliminates the autoregressive decoding bottleneck entirely, yielding massive latency and compute savings for high-frequency agent control flow.
- –Executing directly in the browser via WebGPU proves that responsive decision-making primitives can operate on device with zero hosting cost and complete data privacy.
- –Benchmark data across sizes (from 40.7% on Qwen 0.6B to 84.5% on Qwen 4B) emphasizes that direct logit reliability remains sensitive to model scale and quantization effects.
- –The project demystifies proprietary decision-model marketing, showing that specialized classification behavior can be readily reproduced by developers using standard open-weight architectures.
DISCOVERED
1h ago
2026-09-18
PUBLISHED
4h ago
2026-09-18
RELEVANCE
AUTHOR
ilreb