stepfun-ai/step3

Introduction

Step3 is our cutting-edge multimodal reasoning model—built on a Mixture-of-Experts architecture with 321B total parameters and 38B active. It is designed end-to-end to minimize decoding costs while delivering top-tier performance in vision–language reasoning. Through the co-design of Multi-Matrix Factorization Attention (MFA) and Attention-FFN Disaggregation (AFD), Step3 maintains exceptional efficiency across both flagship and low-end accelerators.

Step3 model card:

| Config | Value | |------------------------|---------| | Number of Layers (Dense layer included)|61| |Number of Dense Layers| 5| | Hidden Dimension | 7168 | | Attention Mechanism | MFA | | Low-rank Query Dimension | 2048 | | Number of Query Heads | 64 | | Head Dimension | 256 | |Number of Experts |48| |Selected Experts per Token|3| |Number of Shared Experts| 1| | Max Context Length | 65536 | | Tokenizer | Deepseek V3 | | Total Parameters (LLM) | 316B | | Activated Params per Token | 38B | | Total Parameters (VLM) | 321B |

Evaluation Results


Deployment

> [!Note] > Step3's API is accessible at https://platform.stepfun.com/, where we offer OpenAI-compatible API for you.

Inference with Hugging Face Transformers

We introduce how to use our model at inference stage using transformers library. It is recommended to use python=3.10, torch>=2.1.0, and transformers=4.54.0 as the development environment.We currently only support bf16 inference, and multi-patch for image preprocessing is supported by default. This behavior is aligned with vllm and sglang.

from transformers import AutoProcessor, AutoModelForCausalLM

key_mapping = {
"^vision_model": "model.vision_model",
r"^model(?!\.(language_model|vision_model))": "model.language_model",
"vit_downsampler": "model.vit_downsampler",
"vit_downsampler2": "model.vit_downsampler2",
"vit_large_projector": "model.vit_large_projector",
}

model_path = "stepfun-ai/step3"

processor = AutoProcessor.from_pretrained(model_path, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(model_path,
device_map="auto", torch_dtype="auto",trust_remote_code=True,
key_mapping=key_mapping)

messages = [
{
"role": "user",
"content": [
{"type": "image", "image": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/bee.jpg"},
{"type": "text", "text": "What's in this picture?"}
]
},
]

inputs = processor.apply_chat_template(
messages, add_generation_prompt=True, tokenize=True,
return_dict=True, return_tensors="pt"
).to(model.device)

generate_ids = model.generate(**inputs, max_new_tokens=32768, do_sample=False)
decoded = processor.decode(generate_ids[0, inputs["input_ids"].shape[-1] :], skip_special_tokens=True)

print(decoded)

Inference with vLLM and SGLang

Our model checkpoints are stored in bf16 and block-fp8 format, you can find it on Huggingface.

Currently, it is recommended to run Step3 on the following inference engines:

• vLLM
• SGLang

Deployment and Request examples for vLLM and SGLang can be found in the [Model Deployment Guide](docs/deploy_guidance.md).

Contact Us

If you have any questions, please reach out at [contact@stepfun.com](mailto:contact@stepfun.com) .

License

Both the code repository and the model weights are released under the [Apache License (Version 2.0)](./LICENSE).

Citation

@misc{step3system,
title={Step-3 is Large yet Affordable: Model-system Co-design for Cost-effective Decoding},
author={StepFun Team},
year={2025},
eprint={2507.19427},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2507.19427},
}

@misc{step3blog,
title={Step3: Cost-Effective Multimodal Intelligence},
author={StepFun Team},
url={https://stepfun.ai/research/step3},
}