STEP3-VL-10B

### TL;DR

STEP3-VL-10B is a lightweight open-source foundation model designed to redefine the trade-off between compact efficiency and frontier-level multimodal intelligence. Despite its compact 10B parameter footprint, STEP3-VL-10B excels in visual perception, complex reasoning, and human-centric alignment.

Key Insights & Metrics

Key Features

• Unified pre-training on 1.2T multimodal tokens
• Integration of a language-aligned Perception Encoder with a Qwen3-8B decoder
• Scaled post-training pipeline with over 1,000 iterations of reinforcement learning
• Implementation of Parallel Coordinated Reasoning (PaCoRe) for scalable perceptual reasoning

Discussion