How does the Step Plan credit system work?

You subscribe to a tier (Mini, Plus, Pro, Max) which provides a monthly pool of Credits. These Credits are consumed by API calls based on model usage and token count. Higher tiers offer more Credits, priority API rates, and better support.

Can I fine-tune these models on my own data?

Yes. StepFun provides open weights for its flagship models, which are explicitly designed to be deployable, customizable, and fine-tunable on your own infrastructure.

Why are some models, like the image generation ones, no longer available?

StepFun periodically retires older or less-aligned models to focus resources on its core, high-performance language and multimodal foundation models (like the Step series).

How do the RPM, TPM, and concurrency rate limits interact?

They are independent limits. You will be blocked if you hit ANY one of them. For example, if your RPM is 20 and TPM is 200K, 20 requests with 100 tokens each will hit the RPM limit first, even if the TPM is not reached.

What does 'native multimodal' mean for Step 3.7 Flash, and why is it better?

It means the model processes text and visual inputs (like images of UIs or charts) in its core architecture. This avoids the extra latency, cost, and potential information loss of using a separate Vision MCP (Model Context Protocol) relay.

How should I calculate costs for Chinese text?

StepFun uses a token-based billing system. For Chinese text, a general rule is that 1 token equals approximately 1.5 to 2 Chinese characters. The exact ratio varies by content. You should monitor your token usage in the dashboard for accurate cost tracking.

Directory/Coding/AI Development/StepFun

StepFun

StepFun is an AI platform offering a suite of proprietary large language models (LLMs) and multimodal models, including the Step series. It provides APIs for developers to build applications ranging from coding assistants to autonomous agents. The platform features models like Step 3.7 Flash, optimized for real-world agent workflows with native multimodal understanding, and Step 3.5 Flash, designed for high-speed reasoning. It also includes a conversational AI product (JieYue AI) and a development studio for building AI applications. The core offering is a stable, high-performance API for integrating these models. The platform supports text generation, code generation, multimodal tasks (understanding images/PDFs), and image generation. It targets developers and companies needing scalable AI capabilities for automation, productivity, and research.

Coding

Visit website

Added on

Sep 19, 2026

Multimodal LLMAgent FrameworkOpen WeightsCredit-based APIDesktop & Mobile AppsChinese-first

Product information

Everything worth knowing about StepFun

The complete picture — from everyday features to the technical detail builders and IT folks dig for.

Core features

What it actually does

Multi-modal AI Models

Step 3.7 Flash is a high-efficiency multimodal model for real agent workflows, natively understanding text, images, and video. It can parse PDFs, tables, charts, and UI screenshots in a single pass.

Language Model Suite

Offers a range of language models including Step 3.5 Flash (fast reasoning), Step 2 (trillion-parameter base), and Step 1 (efficient, low-bit inference) for tasks from content generation to complex reasoning.

Open Platform & API

Provides stable, high-performance Step API for easy integration, allowing developers to build AI applications and connect models, tools, and workflows.

Desktop & Mobile Companion

The 'Step AI' is a customizable AI partner available as desktop (Mac, Windows) and mobile (iOS, Android) applications, plus a web interface.

Technical capabilities

Agent Framework Compatibility

Models feature stable tool calling and deep compatibility with mainstream Agent frameworks, making them suitable for building autonomous agent systems.

Open Weights & Deployment

Flagship models are open-weight, allowing for on-premise deployment, customization, and fine-tuning.

Must watch videos

Let's Run Step-3.5-Flash - SUPER FAST Local AI that Beats GLM & OpenClaw? REVIEW

15:36

Review15:36

Let's Run Step-3.5-Flash - SUPER FAST Local AI that Beats GLM & OpenClaw? REVIEW

One of the six tigers of China have released their latest M4 Max compatible model touting chart topping results at flash speeds. So let's jump in and see how it runs locally. UPDATE: The looping and OpenClaw compatibility issues have now been resolved. TEST SYSTEM Inferencer App v1.9.6: https://inferencer.com Model: https://huggingface.co/inferencerlabs/Step-3.5-Flash-MLX-6.5bit Parallels Virtual Machine: https://vtudio.com/a/?parallels xCreate App: https://xcreate.com 2025 M3 Ultra Mac Studio | 512GB RAM BUY NOW Mac Studio: https://vtudio.com/a/?a=mac+studio MacBook Pro: https://vtudio.com/a/?a=macbook+pro LG C2 42" Monitor: https://vtudio.com/a/?a=lg+c2+42 Recommended NAS Drive: https://vtudio.com/a/?a=qnap+tvs-872xt COMPANION VIDEOS Kimi K2.5 with OpenClaw: https://youtu.be/ExzAiMjT6jg Mixture of Expert Control: https://youtu.be/h4lMWOk_iZk GLM-4.7 Review: https://youtu.be/E-8KJpUFalM Z-Image-Turbo: https://youtu.be/RG5aSqRxAws SPECIAL THANKS Thanks for your support and if you have any suggestions or would like to help us produce more videos, please visit: https://vtudio.com/a/?support CREDITS Six Tigers made locally with https://xcreate.com Links to products often include an affiliate tracking code which allow us to earn fees on purchases you make through them.

Step-3.5 Flash First Look & Testing – An INSANELY Fast Agentic Model!

27:38

Tutorial27:38

Step-3.5 Flash First Look & Testing – An INSANELY Fast Agentic Model!

Timestamps: 00:00 - Intro 00:25 - First Look 01:04 - Technical Look 03:30 - Browser OS Test 06:51 - Browser OS Test V2 08:00 - 3D Printer Simulation Test 10:49 - Python 3D FPS Test 12:54 - Flight Combat Simulator Test 14:14 - Flight Combat Simulator V2/V3 16:24 - Highway Racing Game Test 17:01 - GLM-4.7 & MiniMax Comparison Test 18:05 - Highway Racing Game V2/V3 Test 19:20 - European Design Firm Website Test 20:40 - Swedish Design Firm Website Test 22:04 - Bonus Website Test 23:00 - Finance Website Test 24:28 - Model Summary 24:51 - DGX Spark Attempt & Failure FYI 25:27 - Results Overview & Closing Thoughts AI Integration & Consulting https://bijanbowen.com Join The Discord: https://discord.gg/hfaR2exy7S In this video, we take a first look at Step-3.5 Flash, a newly released model from StepFun-AI designed with speed and agentic workflows in mind. This model uses a sparse MoE architecture at just under 200B parameters, with a strong emphasis on extremely fast inference and practical agentic performance. We begin with a technical overview of the model and its design goals, including its focus on running efficiently on unified-memory systems such as DGX Spark and Strix Halo–class hardware. From there, we move into a wide range of hands-on tests to evaluate how well Step-3.5 Flash performs in real-world scenarios.

Step 3.5 Flash : This Open-Source AI Model BEATS GLM 4.7 and Deepseek 3.2

11:04

Tutorial11:04

Step 3.5 Flash : This Open-Source AI Model BEATS GLM 4.7 and Deepseek 3.2

Step 3.5 Flash is a powerful open-source large language model (LLM) built for high-speed reasoning, AI agents, and real-time applications. With a Sparse Mixture-of-Experts (MoE) architecture, massive 196B total parameters, and only 11B active per token, this model delivers top-tier intelligence at ultra-low compute cost. =================================================== https://huggingface.co/stepfun-ai/Step-3.5-Flash https://openrouter.ai/stepfun/step-3.5-flash:free https://stepfun.ai =================================================== #aiagents #OpenSourceAI #ai #aicoding #LLM #AIModel #MachineLearning #NaturalLanguageProcessing #AIResearch #ainews #HuggingFace #OpenRouter #AICommunity #TechInnovation #AIDevelopment #SoftwareEngineering Disclaimer: The content in this video is for informational and educational purposes only. All opinions expressed are my own. I am not a licensed professional, and this video should not be considered professional advice. Performance benchmarks are based on specific tests and may not reflect all use cases. Always do your own research and consult with a qualified expert where necessary. Use the information provided at your own risk. Some links or products mentioned may be affiliate links, which means I may earn a small commission at no extra cost to you. Thank you for your support!

Who uses it

Real use cases, no hype

Agent Workflow Automation

Build and deploy autonomous agents for coding, data analysis, and task execution with reliable tool calling.

Enterprise Multimodal Document Processing

Extract and reason over text, tables, and charts from PDFs, screenshots, and UI layouts in one pass.

AI-Powered Code Generation & Refactoring

Leverage flagship coding models to generate, explain, and optimize code within IDEs or standalone tools.

Real-time Visual Customer Support

Deploy chatbots that can interpret user screenshots or photos to provide visual-guided troubleshooting.

Academic & Research Analysis

Use long-context models to synthesize insights from dense papers, theses, and multi-source research data.

Cross-Modal Creative Generation

Generate images from text prompts or create descriptive text from visual concepts using integrated models.

Scalable Internal Tools via API

Integrate stable, high-performance text models into internal apps for search, summarization, and classification.

What's great

Flagship multimodal model (Step 3.7 Flash) with native visual understanding for UIs, PDFs, and charts.
Full open weights for flagship models enabling self-hosting, customization, and fine-tuning.
Deep compatibility with major Agent frameworks and stable tool calling for real workflows.
Competitive credit-based pricing tiers (Step Plan) with priority support for higher tiers.

Technical strengths

Efficient single-card inference for older models (Step 1) with low-bit, long-text capabilities.
Native multimodal processing eliminates need for separate Vision MCP, reducing latency.
Hybrid token/RPM/concurrent rate limits provide granular control for high-throughput systems.

Where it falls short

Ecosystem is China-centric; English documentation and community support may lag.
Tiered pricing can be complex; high-volume needs require enterprise contact for bulk deals.
Newer model versions (e.g., Step 3.5) released rapidly, requiring frequent migration.
Some model offerings (e.g., image generation) have been sunsetted, causing platform fragmentation.

Technical limitations

API rate limits (RPM, TPM, concurrent) are strict and can trigger unexpectedly.
Token counting for Chinese is non-intuitive (~1.5-2 chars per token), complicating cost prediction.
Flash-tier models are optimized for speed, potentially sacrificing complex reasoning depth.

For technical folks

The deep-dive specs

FAQ

Includes technical Q&A

Step 3.7 Flash is a multimodal, efficiency-optimized model for real-world agent workflows, with native visual understanding. Step 2 is a larger, trillion-parameter language model focused on deep reasoning and instruction following.

Traffic Insights

Monthly Visits

—

November 2026

Growth Rate

vs last month

Dominance

in Coding

Top Country

—

Top Countries Breakdown

—

Growth Trend

Stable

Traffic has remained stable.

Pricing overview

Freemium Pricing

StepFun Pricing

StepFun uses a freemium pricing structure. Real-time rates, custom enterprise quotas, and plan details are managed directly on their official platform.

Check Pricing on StepFun

Reviews