TypeSafe AI Unveils 'Jev', a Structured Decision Model 200x Faster Than LLMs

TypeSafe AI has released 'Jev', a System One model class designed to perform fast, structured decisions instead of text generation. The company claims that for classification tasks, the model offers inference speeds up to 200x faster and costs 400x lower compared to comparable LLMs.

Jev is trained using the Reinforcement Learning for Calibrated Decisions (RLCD) technique and returns typed answers with probabilities for given input states and questions. It allows code to utilize results to determine an agent's next action without full LLM calls. Question types consist of three categories: Choice (option selection), Score (ordered level assessment), and Noul (yes/no questions).

LangChain Integration and Use Cases

Jev is designed as a complement, not a replacement, for LLMs. While LLMs handle open-ended reasoning and generation, Jev processes fast structural decisions. To integrate with LangChain, install the langchain-typesafe package and use TypeSafeClassifier.

Key use cases include:

• Model Routing: Use ModelRouterMiddleware to automatically select between 'fast' and 'powerful' models based on request complexity.
• Auto Mode: Use AutoModeMiddleware to classify actions with Jev and apply guardrails before executing risky behaviors like bash tools. This functionality, previously limited to closed-source harnesses, is now available for all agents.

Kyle Jeong of Browserbase utilized Jev to run browser-using agents for less than 1 cent, while Jarrod Watts and Ryan Vogel applied it to live trading agents and large-scale email classification, respectively.