Thomas Wolf says open Decision 3.0 models are state-of-the-art after apple-harvesting test

Thomas Wolf of Hugging Face says open decision models have reached state-of-the-art performance.

In an X post, he cites a test run by Victor Mustar in which Decision 3.0 was the top performer, collecting 7 apples in 38 moves with no damage.

The model reads screenshots, game stats and its last four moves, then chooses up, down, left or right on each turn. The reported result comes from a single apple-harvesting test and does not include other benchmarks.