Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

STEP3-VL-10B Technical Report

Ailin Huang, Chengyuan Yao, Chunrui Han, Fanqi Wan, Hangyu Guo, Haoran Lv, Hongyu Zhou, Jia Wang, Jian Zhou, Jianjian Sun, Jingcheng Hu, Kangheng Lin, Liang Zhao, Mitt Huang, Song Yuan, Wenwen Qu, Xiangfeng Wang, Yanlin Lai, Yingxiu Zhao, Yinmin Zhang, Yukang Shi, Yuyang Chen, Zejia Weng, Ziyang Meng, Ang Li, Aobo Kong, Bo Dong, Changyi Wan, David Wang, Di Qi, Dingming Li, En Yu, Guopeng Li, Haiquan Yin, Han Zhou, Hanshan Zhang, Haolong Yan, Hebin Zhou, Hongbo Peng, Jiaran Zhang, Jiashu Lv, Jiayi Fu, Jie Cheng, Jie Zhou, Jisheng Yin, Jingjing Xie, Jingwei Wu, Jun Zhang, Junfeng Liu, Kaijun Tan, Kaiwen Yan, Liangyu Chen, Lina Chen, Mingliang Li, Qian Zhao, Quan Sun, Shaoliang Pang, Shengjie Fan, Shijie Shang, Siyuan Zhang, Tianhao You, Wei Ji, Wuxun Xie, Xiaobo Yang, Xiaojie Hou, Xiaoran Jiao, Xiaoxiao Ren, Xiangwen Kong, Xin Huang, Xin Wu, Xing Chen, Xinran Wang, Xuelin Zhang, Yana Wei, Yang Li, Yanming Xu, Yeqing Shen, Yuang Peng, Yue Peng, Yu Zhou, Yusheng Li, Yuxiang Yang, Yuyang Zhang, Zhe Xie, Zhewei Huang, Zhenyi Lu, Zhimin Fan, Zihui Cheng, Daxin Jiang, Qi Han, Xiangyu Zhang, Yibo Zhu, Zheng Ge

TL;DR

STEP3-VL-10B introduces a compact 10B-parameter foundation model that closes the gap to frontier multimodal intelligence through two main innovations: a unified unfrozen pre-training on $1.2T$ multimodal tokens that couples a language-optimized Perception Encoder with a $Qwen3-8B$ decoder, and a scaled post-training pipeline featuring PaCoRe-assisted reinforcement learning with over $1{,}000$ iterations. This design yields strong multimodal performance and robust reasoning while maintaining efficiency, enabling comparisons with models up to $10\times$–$20\times$ larger and even some proprietary frontiers. The post-training regime combines two-stage supervised finetuning (SFT) with reinforcement learning (PPO/GAE) under verifiable and non-verifiable reward structures, and scales reasoning via Latent parallel coordination to synthesize diverse perceptual hypotheses. Extensive evaluations across 60+ benchmarks show STEP3-VL-10B achieving top-tier results in both multimodal and text-centric tasks, with notable gains in GUI grounding, OCR, math reasoning, and human-alignment, underscoring its potential as an open, efficient baseline for large-scale multimodal intelligence.

Abstract

We present STEP3-VL-10B, a lightweight open-source foundation model designed to redefine the trade-off between compact efficiency and frontier-level multimodal intelligence. STEP3-VL-10B is realized through two strategic shifts: first, a unified, fully unfrozen pre-training strategy on 1.2T multimodal tokens that integrates a language-aligned Perception Encoder with a Qwen3-8B decoder to establish intrinsic vision-language synergy; and second, a scaled post-training pipeline featuring over 1k iterations of reinforcement learning. Crucially, we implement Parallel Coordinated Reasoning (PaCoRe) to scale test-time compute, allocating resources to scalable perceptual reasoning that explores and synthesizes diverse visual hypotheses. Consequently, despite its compact 10B footprint, STEP3-VL-10B rivals or surpasses models 10$\times$-20$\times$ larger (e.g., GLM-4.6V-106B, Qwen3-VL-235B) and top-tier proprietary flagships like Gemini 2.5 Pro and Seed-1.5-VL. Delivering best-in-class performance, it records 92.2% on MMBench and 80.11% on MMMU, while excelling in complex reasoning with 94.43% on AIME2025 and 75.95% on MathVision. We release the full model suite to provide the community with a powerful, efficient, and reproducible baseline.

STEP3-VL-10B Technical Report

TL;DR

STEP3-VL-10B introduces a compact 10B-parameter foundation model that closes the gap to frontier multimodal intelligence through two main innovations: a unified unfrozen pre-training on multimodal tokens that couples a language-optimized Perception Encoder with a decoder, and a scaled post-training pipeline featuring PaCoRe-assisted reinforcement learning with over iterations. This design yields strong multimodal performance and robust reasoning while maintaining efficiency, enabling comparisons with models up to – larger and even some proprietary frontiers. The post-training regime combines two-stage supervised finetuning (SFT) with reinforcement learning (PPO/GAE) under verifiable and non-verifiable reward structures, and scales reasoning via Latent parallel coordination to synthesize diverse perceptual hypotheses. Extensive evaluations across 60+ benchmarks show STEP3-VL-10B achieving top-tier results in both multimodal and text-centric tasks, with notable gains in GUI grounding, OCR, math reasoning, and human-alignment, underscoring its potential as an open, efficient baseline for large-scale multimodal intelligence.

Abstract

We present STEP3-VL-10B, a lightweight open-source foundation model designed to redefine the trade-off between compact efficiency and frontier-level multimodal intelligence. STEP3-VL-10B is realized through two strategic shifts: first, a unified, fully unfrozen pre-training strategy on 1.2T multimodal tokens that integrates a language-aligned Perception Encoder with a Qwen3-8B decoder to establish intrinsic vision-language synergy; and second, a scaled post-training pipeline featuring over 1k iterations of reinforcement learning. Crucially, we implement Parallel Coordinated Reasoning (PaCoRe) to scale test-time compute, allocating resources to scalable perceptual reasoning that explores and synthesizes diverse visual hypotheses. Consequently, despite its compact 10B footprint, STEP3-VL-10B rivals or surpasses models 10-20 larger (e.g., GLM-4.6V-106B, Qwen3-VL-235B) and top-tier proprietary flagships like Gemini 2.5 Pro and Seed-1.5-VL. Delivering best-in-class performance, it records 92.2% on MMBench and 80.11% on MMMU, while excelling in complex reasoning with 94.43% on AIME2025 and 75.95% on MathVision. We release the full model suite to provide the community with a powerful, efficient, and reproducible baseline.
Paper Structure (114 sections, 3 equations, 7 figures, 8 tables)

This paper contains 114 sections, 3 equations, 7 figures, 8 tables.

Table of Contents

  1. Introduction
  2. Pre-train
  3. Architecture
  4. Data Construction
  5. Knowledge.
  6. Education.
  7. Optical Character Recognition (OCR).
  8. Grounding & Counting.
  9. Visual Question Answering (VQA).
  10. Graphical User Interface (GUI).
  11. Training Recipe
  12. Post-Train
  13. Supervised Finetuning
  14. Data Construction.
  15. Two-Stage SFT Strategy.
  16. ...and 99 more sections

Figures (7)

  • Figure 1: Performance comparison of Step3-VL-10B against state-of-the-art multimodal foundation models. With PaCoRe (Parallel Coordinated Reasoning hu2025pacore, Step3-VL-10B scales test-time compute to bridge the perception and reasoning performance gap with 100B+ parameter models.
  • Figure 2: RLVR dynamics. While the reward continuously increases without saturating (right), the average rollout tokens decrease towards the starting level after an initial rise (left).
  • Figure 3: Trends of representative multimodal reasoning and perception metrics during RLVR. Evaluated every 100 iterations, performance mirrors the reward dynamics: rapid initial growth followed by steady improvement.
  • Figure 4: Morse Code Reference
  • Figure 5: Screenshot of the Compiler
  • ...and 2 more figures