A new preprint introduces CHSM8, a chess transformer with 210 million parameters that generates moves based on free-text descriptions of a desired playing style. Rather than simply maximizing winning chances, the model is designed to follow stylistic instructions such as "go for the Sicilian Defense" or "solid, safe, risk-free chess." The abstract also hints at persona-based prompts, with an example beginning "You are Magnus Carlsen in his..." (the text is truncated in the announcement).
The model appears to treat style as a conditioning signal, allowing the same underlying network to produce different behaviors depending on the prompt. This could make chess AI more useful for training, entertainment, or analysis, where users want to specify not just strength but temperament and opening preferences. The work also illustrates how transformer architectures, originally developed for language, can be adapted to structured game domains with explicit textual control.
Because only one source was available, there are no conflicting accounts to compare. The findings are preliminary and have not yet been peer-reviewed, but they point toward a future where game agents respond to high-level human instructions rather than only low-level heuristics.