News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
imgStep3

Step3

Multimodal reasoning model with 321 billion parameters

Step3 is a multimodal reasoning model developed by StepFun AI. It uses a mixture of experts architecture with 321 billion total parameters and 38 billion active per token. The model handles text and images, aiming for efficient decoding on various hardware accelerators.

Share on WhatsApp

  • 321 billion total parameters with 38 billion active
  • Supports text and image inputs for reasoning
  • Released under Apache License 2.0
  • Uses Multi-Matrix Factorization Attention
  • Context length of 65,536 tokens

What it is

Step3 is a large language model designed for multimodal reasoning tasks. StepFun AI built the system to process both text and visual inputs. The model contains 321 billion parameters in its vision-language configuration. It activates only 38 billion parameters for each token during inference. This design aims to reduce computational costs while maintaining high performance. The model supports a context window of 65,536 tokens. It is intended for use on both high-end and lower-end hardware accelerators.

How it came about

StepFun AI released Step3 as part of its research into cost-effective decoding. The company published technical reports detailing the model-system co-design. These documents explain how the architecture minimises decoding expenses. The release includes model weights and code under an open licence. The team focused on balancing flagship performance with affordability for broader deployment.

Inside the model

The model uses a mixture of experts architecture with 61 layers. Five of these layers are dense, while the rest use sparse activation. It employs Multi-Matrix Factorization Attention to improve efficiency. The design also features Attention-FFN Disaggregation. Each token selects three experts from the available pool, with one shared expert. The hidden dimension is 7,168, and it uses 64 query heads. The tokenizer is based on Deepseek V3. Checkpoints are stored in bf16 and block-fp8 formats to support efficient inference.

What it can do

Step3 is built for vision-language reasoning tasks. It processes images and text together to answer questions about visual content. The system claims top-tier performance in this domain. It is designed to maintain efficiency across different types of hardware. The model can generate up to 32,768 new tokens in a single response. Evaluation results are referenced in the company blogs, though specific scores are not detailed in the provided text.

How to use it

The model weights and code are available under the Apache License 2.0. Users can download them from Hugging Face. StepFun AI also provides an API that is compatible with OpenAI standards. Inference can be run using libraries such as Transformers, vLLM, or SGLang. The company recommends using Python 3.10 and PyTorch 2.1.0 or higher for development. Only bf16 inference is currently supported.

Questions people ask about Step3

What is the parameter count of Step3?

Step3 has 321 billion total parameters in its vision-language model configuration. However, it uses a mixture of experts design, so only 38 billion parameters are active for each token processed.

Can I use Step3 for free?

Yes, the model weights and code are released under the Apache License 2.0. You can download them from Hugging Face and run them locally, or use the company's API which may have its own pricing.

What types of data does Step3 accept?

Step3 is a multimodal model that accepts both text and images. It is designed for vision-language reasoning, allowing users to ask questions about pictures or analyse visual content alongside text.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Step3 yet, so the facts on this page come from the sources listed above: huggingface.co/stepfun-ai/step3. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.