Tokamak experiments record waveforms, profiles, fields and scalar measurements that support plasma reconstruction, evolution forecasting, control‑signal recovery and disruption‑risk estimation. Existing models are usually built for a single task with a fixed input‑output set and rely on a predefined subset of diagnostics, missing the opportunity to exploit the full multimodal observation for global cross‑signal insight. To address this, we extend the TokaMark benchmark with two new task families—actuator action inversion and disruption prediction—resulting in 20 MAST tasks organized into six groups, each with explicit schemas and evaluation protocols. We then introduce GenToka, a generative tokamak foundation model that jointly generates schema‑selected targets via a shared conditional diffusion Transformer, while signal‑specific codecs preserve native time‑series, profile, field and scalar formats. The model undergoes foundation pre‑training followed by task‑specific adaptation, transferring cross‑signal structure across all six groups. Evaluation shows that GenToka reduces macro‑average normalized root‑mean‑square error by 4.7% relative to TokaMind on Groups 1‑4 and by 17.7% on Group 5, and improves macro‑average window‑level F1 by 8.4% on Group 6 using a common validation‑selected threshold. GenToka offers a reusable, generative, multimodal pathway for experimental tokamaks, linking plasma‑state reconstruction, actuator inference and disruption warning from shared measurements, thereby supporting more integrated, data‑driven fusion plasma prediction and control research.\n\nReview: The study convincingly demonstrates the practical benefits of a cross‑task, cross‑modal foundation model for fusion experiments, laying a solid technical foundation for unified experimental data platforms.