#22StepFun

Step 3.7 Flash

An open-weights vision-language mixture-of-experts model tuned for coding agents, search and multimodal workflows. It pairs a 196B language backbone with a 1.8B vision encoder, activates about 11B parameters per token, and is released under Apache 2.0. It is also available through StepFun's API, OpenRouter and NVIDIA NIM.

• 198B total (196B + 1.8B ViT) / ~11B active
• 256K context; text and image input
• SWE-Bench Pro 56.3%, BrowseComp 75.8% (lab)
• Apache 2.0; FP8, NVFP4 and GGUF builds

AccessOpen weights (Apache 2.0) on Hugging Face/ModelScope; StepFun API, web and mobile apps; OpenRouter; NVIDIA NIM

Price$0.20 input / $1.15 output per 1M tokens

Step 3.5 Flash

A fast open-weights reasoning model aimed at agents. Its sparse mixture-of-experts design with multi-token prediction runs at 100-350 tokens per second. StepFun reports maths and coding scores close to those of much larger models, and also released base and mid-training checkpoints.

• 196B total / ~11B active parameters
• 256K context (3:1 sliding-window attention)
• SWE-bench Verified 74.4%, AIME 2025 97.3 (lab)
• Apache 2.0; base and midtrain checkpoints

AccessOpen weights (Apache 2.0) on GitHub, Hugging Face, ModelScope; StepFun API; OpenRouter

Price$0.10 input / $0.30 output per 1M tokens

Step-Audio-R1.1

An open 33B audio reasoning model that reasons over sound directly rather than over transcripts. It can 'think while speaking', which keeps latency low in spoken dialogue. It is available as Apache 2.0 weights and through StepFun's Audio Studio API.

• 33B parameters, Apache 2.0
• Chain-of-thought grounded in audio, not text
• 'Mind-paced speaking': reasons during speech output
• Real-time spoken dialogue

AccessOpen weights on Hugging Face (Apache 2.0); StepFun Audio Studio API

Step3-VL-10B

A compact 10B open vision-language model that combines a 1.8B perception encoder with a Qwen3-8B decoder. StepFun targets STEM reasoning, OCR and document analysis, GUI interaction and spatial reasoning, with an optional parallel reasoning mode (PaCoRe) for harder problems. It needs about 24GB of GPU memory.

• 10B parameters (1.8B encoder + Qwen3-8B decoder)
• MMMU 80.11, AIME 2025 94.43 with PaCoRe (lab)
• OCR, documents, GUI and spatial tasks
• Apache 2.0; FP8 build

AccessOpen weights on Hugging Face (Apache 2.0), incl. FP8; Hugging Face Space demo; vLLM

NextStep-1.1

StepFun's open 15B autoregressive image generator, which works with continuous tokens instead of diffusion. Version 1.1 adds flow-based reinforcement-learning post-training to improve texture and cut visual artifacts. It also fixes the image failures and training instability of the original NextStep-1 model.

• 15B parameters, Apache 2.0
• Autoregressive text-to-image with continuous tokens
• RL post-training for fidelity, fewer artifacts
• 512x512 default output

AccessOpen weights on Hugging Face (Apache 2.0)

GELab-Zero-4B-preview

A 4B GUI agent model that runs locally and automates Android phones: it reads the screen, then taps, types and swipes through multi-step tasks in apps it has not seen before. StepFun released it with its full agent infrastructure and an AndroidDaily benchmark.

• 4B parameters; runs on consumer hardware
• Zero-shot operation across unseen Android apps
• AndroidDaily: 75.86% end-to-end success (lab)
• Multi-device management via MCP server

AccessOpen weights on Hugging Face and ModelScope; agent stack on GitHub

Step-Audio-R1

An open 33B audio-language model that StepFun calls the first to make chain-of-thought reasoning work for audio. It grounds its reasoning in acoustic properties instead of transcripts, and handles speech, environmental sound and music. StepFun says it beats Gemini 2.5 Pro on audio reasoning benchmarks.

• 33B parameters, Apache 2.0
• Audio chain-of-thought with test-time scaling
• Speech, environmental sounds and music
• Beats Gemini 2.5 Pro on audio benchmarks (lab)

AccessOpen weights on Hugging Face (Apache 2.0)