Recent testing of multimodal agents has shown an ability to process and react to live visual data in high-speed gaming environments. An agent using the SGLang framework defeated a Level 9 computer opponent in Super Smash Bros Melee, playing as the character Fox on the Final Destination map without items. The victory, which included a 2 stock of the CPU player, was executed in real time through an API.

This development builds on a previous success where a Qwen3.8-27B model beat the Pokémon FireRed champion with responses in under 100 ms. SGLang utilizes a /v1/decisions endpoint to turn large models into classification and scoring tools, alongside a /v1/systemone endpoint for TypeSafe SDK integration. Some observers of earlier tests noted that the majority of the model's critical tactical choices were poor, suggesting the results are primarily a achievement in systems integration.