DreamChipper
7 posts
🚀 Meet STEP3-VL-10B—it delivers SOTA-level visual perception, complex reasoning, and human-aligned intelligence—redefining efficiency in open multimodal AI. ✅ Beats or matches models 10–20× larger (like GLM-4.6V, Qwen3-VL, even Gemini 2.5 Pro) ✅ Achieves SOTA on MMMU, MathVision, OCRBench, ScreenSpot, and more ✅ Trained on 1.2T tokens + 1,400+ RL rounds (RLHF + RLVR) ✅ Supports PaCoRe: parallel collaborative reasoning (128K context!) Despite its compact size, it leads the <10B class in: • STEM reasoning (94.43% on AIME 2025 w/ PaCoRe!) • Visual perception (92.05 on MMBench) • GUI understanding & OCR • Spatial reasoning for embodied AI Built with: • PE-lang visual encoder (1.8B) • Qwen3-8B decoder • Multi-crop high-res input (728×728 global + locals) 📥 Base & Chat versions on ModelScope: Base: modelscope.cn/models/stepfun… Chat: modelscope.cn/models/stepfun…
