A developer has released jevlike, an open-source starter model designed to perform tasks similar to TypeSafe's "Jev" model. The model takes a piece of text and a list of N text options, returning a probability score for each option in a single pass. This approach avoids the word-by-word generation used by standard decoders, offering significant efficiency gains.
Open-source implementation of one-pass "Jev-like" model released
In local experiments, the one-pass scorer reached approximately 98% accuracy on synthetic menus. When tested against Wikispeedia next-click data using a frozen Qwen2.5-0.5B encoder, the model achieved 26% accuracy, compared to 8% for random-encoder controls. At eight options, the one-pass method was approximately 100 times faster than a small decoder forced to generate 400 tokens.
The model functions by using each option as a query vector to assign attention weights to context tokens. This process creates a context vector for each option, which is then scored via a dot product. Users can train the model from scratch or use frozen pretrained encoders from Hugging Face, such as Qwen2.5-0.5B. The repository includes examples for gaming tasks, such as controlling characters in Doom and Chess, demonstrating the model's ability to score decisions from visual or text-based inputs.
Sources
- Reverse-engineered Jev-like model （Hacker News Frontpage, 2026-09-16）
- TypeSafe公式ブログ