Step3-VL-10B

A multimodal version of Qwen3 with a rather beefy 1.7B vision encoder and 1.2T tokens of continued, multimodal training, followed by an RL stage.