AI · AI model
stepfun-ai/Step3-VL-10B
Open 10B multimodal model with parallel reasoning competitive with frontier systems
STEP3-VL-10B is an open-source multimodal foundation model that pairs a 1.8B visual encoder with a Qwen3-8B language decoder. Using Parallel Coordinated Reasoning (PaCoRe), it aggregates multiple visual exploration paths to achieve reasoning, OCR, and GUI grounding benchmark scores that compete with models ten to twenty times its size. It runs on single 24GB GPUs using vLLM, SGLang, or standard Transformers.
What it's for
- Extract structured text and tables from dense documents, screenshots, and handwritten notes
- Power GUI automation and agentic workflows using UI element grounding
- Solve complex visual STEM, math, and geometry problems on local hardware
Features
- 10B parameter architecture pairing PE-lang visual encoder with a Qwen3-8B decoder
- Parallel Coordinated Reasoning (PaCoRe) aggregating 16 parallel rollouts up to 128K context
- Sequential Reasoning (SeRe) standard Chain-of-Thought mode up to 64K context
- Trained on 1.2T multimodal tokens with extensive RLVR and RLHF post-training
- Native support for vLLM and SGLang deployment with OpenAI-compatible APIs
- Available in full bf16 weights and FP8 quantized formats
- High performance in visual perception, GUI grounding, spatial understanding, and OCR
Pricing
Open source. Free and open source under Apache 2.0 license
Details
- Type
- AI model
- Category
- AI
- Topic
- Multimodal Models
- Open source
- Yes
- Platforms
- Linux, Self-hosted, Code library, API
- For
- AI researchers, machine learning engineers, and developers building local vision agents
- Alternative to
- Qwen2.5-VL-7B, GLM-4V, Gemini 2.5 Pro
- Tags
- vlm, vision language, reasoning, ocr, open weights, huggingface
Why it was saved
- @JulianGoldieSEO: Covers OCR (94.2% benchmark, near-perfect document/table/handwriting extraction), GUI and spatial understanding, STEM/math reasoning, text in 20+ languages; learned vision+language together on 1.2 trillion mixed tokens; every query spins up 16 parallel reasoning streams that self-correct and merge into a verified answer, claimed to reduce hallucinations; scores 92.2% on multimodal benchmarks; optimized for M3/M4 and RTX 40-series, open weights on Hugging Face, works offline.
Related
Open-source vision-language model converting complex documents into structured Markdown using reasoning.
Open-source vision-language model converting complex document images to structured Markdown
Open-source vision-language model for parsing multi-page documents in one pass
Open-source 4B flow-matching model generating textured 3D assets from single images.
Open full-duplex speech-to-speech conversational AI model by NVIDIA
Open-source 1.6B text-to-speech model generating multi-speaker dialogue in one pass.
Updated