Skip to main content
Rase AI Intelligence Newsletter

Models & assistants Model or AI service

Step-3 is a downloadable vision-language model designed around inference cost

StepFun’s Step-3 processes visual and text inputs. Its report describes a 321B-parameter model designed together with the system that runs it. [1][2]

Model access
StepFun makes Step-3 weights available for download and lists access options. [2]
License
Apache License 2.0 for the published Step-3 weights and project code. [2]

What the model does

StepFun’s July 2025 report describes Step-3 as a mixture-of-experts model that activates about 38B parameters per token. The project provides weights and separately documents serving requirements. [1][2]

Why it matters

Separating the total size from the part used at each step helps explain why a system can run fewer calculations per token while still needing room for a large checkpoint. [1][2]

Who can use it

Model researchers and inference teams can evaluate the checkpoint for visual question answering and other tasks that combine text and images. [1][2]

Key dates

  1. Step-3 report first submitted [1]

Ways to use it

Test visual questions on example material and measure the time it takes to return a useful answer. Compare with another model in the same serving setup so hardware differences do not dominate the results. [1][2]

Model limitations

The report’s performance measurements depend on a specific hardware and serving configuration. The full checkpoint is far larger than the number of parameters used for one token. [1][2]

Sources and verification

Catalog evidence checked How we verify records

  1. Primary source StepFun’s initial Step-3 technical report and submission date
  2. Primary source stepfun-ai/step3
  3. Other source Step-3 is Large yet Affordable: Model-system Co-design for Cost-effective Decoding

Related profiles

Models & assistants

GPT-6.2 Astra and Sol

OpenAI

Astra targets demanding projects. Sol targets everyday tasks. Both support text and image input in the API.

High confidenceEvidence checked

Models & assistants

MAI Code 2 and Code 2 Flash

Microsoft AI

Code 2 targets complex software work. Flash targets frequent edits, explanations and routine fixes.

High confidenceEvidence checked

Models & assistants

MAI Realtime 1

Microsoft AI

Users can try a MAI Labs demo. Developers can register for a limited Foundry preview.

High confidenceEvidence checked