Skywork-R1V4

Despite recent advances in multimodal agentic systems, most approaches treat image manipulation and web search as isolated skills, rely on expensive reinforcement learning, and lack planning grounded in real tool-execution traces. These models often fall into one of two camps: (1) reasoning-centric systems (e.g., o1 or earlier Skywork-R1V variants) that perform deep but static "brain-in-a-vat" inference without real-world interaction, or (2) agentic RL frameworks that act in the world but struggle with sustained, reflective planning.

We present Skywork-R1V4, an multimodal agentic model that unifies these dichotomies. R1V4 integrates multimodal planning, active image manipulation ("thinking with images"), deep multimodal search, and—critically—interleaved reasoning that dynamically alternates between visual operations and external knowledge retrieval. Trained solely via supervised fine-tuning on fewer than 30K high-quality, planning-execution-consistent trajectories—and validated through step-wise consistency filtering—R1V4 achieves state-of-the-art results: 66.1 on MMSearch, 38.4 on BrowseComp-VL, 88.0 on V Star and 71.4 on MME-RealWorld, outperforming Gemini 2.5 Flash. Remarkably, this was accomplished in under 12 hours on 64 A800 GPUs at low cost—without any reinforcement learning.