LLaVA-OneVision-1.5 is a fully open-source framework for democratized multimodal training from EvolvingLMMs Lab. It operates on native-resolution images using a three-stage training pipeline with an 85M concept-balanced pretraining dataset and 22M instruction dataset. The complete framework trains state-of-the-art 4B and 8B multimodal models within a $16,000 compute budget, surpassing Qwen2.5-VL across multiple benchmarks. A reinforcement learning extension was also released in late 2025.