Non‑LLM model Jev beats Pokémon Red in a week

TypeSafe AI's Jev, a decision model built for speed rather than language reasoning, beat Pokémon Red in under a week. It reached the Hall of Fame on September 23, Tom's Hardware reported.

Jev outputs a structured, probabilistic choice from a short list of options in 50 to 150 milliseconds per decision, rather than reasoning in natural language like a chatbot, according to the project's public game log. That is far faster than the large language models that have taken weeks or months to finish the same game.

Claude Opus 5 played a supporting role, watching Jev's game log and rewriting the wording and structure of its choices when it got stuck, Tom's Hardware reported. Jev made the moves. Claude wrote the rules Jev chose from.

The run is documented on GitHub and was streamed live on YouTube as Jev fought through the Elite Four and the champion battle, according to the project's public pages. TypeSafe AI markets Jev as a fast, structured alternative to routing tasks through a full chatbot.