TypeSafe Jev Played Chess — And Landed Next to Reasoning Models

TypeSafe's Jev is a unique model that differs from standard chat models. Instead of engaging in conversations, Jev takes typed questions in the form of Choice, Score, or Noul and generates labels with probabilities, akin to a classifier. The model can handle free text inputs and allows users to choose the labels each time. To test its capabilities, TypeSafe utilized the LLM Chess harness, which involves playing chess through an agentic setting with a multi-turn protocol.

Jev follows this protocol, receiving current board information, legal moves, and making moves using UCI strings.

Unlike other models, TypeSafe is openly against benchmarking and prefers internal snapshots over traditional public-eval scoreboards. By putting Jev on the LLM Chess leaderboard, the model was placed next to other reasoning models, such as Qwen3.6-27b and O4-mini-medium, with an Elo rating of around 243. The Elo rating serves as a proxy for reasoning capability, placing Jev in the same band as mid-pack reasoning/small models.

The game duration for Jev was consistently around 36 seconds per game, and the cost per game was approximately $0.0015, making it an extremely fast and inexpensive option.

In a test against the Dragon model, Jev achieved a 50% draw rate at levels L1, L2, and L3, similar to other strong small models. This result suggests that despite being a "System One" model, Jev can still hold its own against more advanced reasoning models in terms of Elo rating and cost-effectiveness. The model's effectiveness lies in its ability to choose from a list of options, such as chess moves, rather than generating lengthy English text.

This makes Jev particularly useful for tasks that require structured decision-making, such as routing or decision steps in code, rather than generating lengthy paragraphs.