StepFun Step 3.5 Flash Pre-training/Mid-training/Training Framework All Open Source!
Published · Mar 4 · Wed Source · 阶跃星辰 (CN)

StepFun Step 3.5 Flash Pre-training/Mid-training/Training Framework All Open Source!

StepFun announces that the Step 3.5 Flash pre-training, mid-training, and training framework are all open source, including Base weights, Midtrain weights, and the Steptron training framework. This model is StepFun's most capable open-source foundation model to date, featuring powerful reasoning and Agent intelligence, designed specifically for Agents.

KeywordsAgentStepFunStepFlashPre-trainingMid-trainingTrainingFramework

step-3.5-flash is StepFun's flagship language reasoning model. It possesses top-tier reasoning capabilities and fast, reliable execution abilities, capable of completing complex task decomposition and planning. It can quickly and reliably call tools to execute tasks, handling various complex tasks such as logical reasoning, mathematics, software engineering, and deep research. Based on a sparse MoE architecture with 196B total parameters / 11B active parameters.

Key Information

Model Type

Sparse MoE Architecture

196B Total Parameters / 11B Active Parameters

Context Length

256K tokens

Scenario Positioning

High-speed reasoning + Tool calling

Optimized for Agent and code tasks

Core Capabilities

🚀 High-speed Reasoning

Sparse MoE architecture brings high throughput and low latency, suitable for real-time agent workflows and high-frequency calling scenarios.

🛠️ Tool Calling

Reliable tools / tool_choice calling capabilities, supporting multi-step task decomposition and plan execution.

🧠 Complex Reasoning

Capable of handling complex tasks such as logical reasoning, mathematics, software engineering, and deep research, serving as a stable foundation for Agent long-chain reasoning.

Model Variants

step-3.5-flash

Base version: General reasoning and tool calling, suitable for most Agent and complex task scenarios.

step-3.5-flash-2603

Agent Optimized version: Optimized for high-frequency Agent scenarios based on step-3.5-flash, with improved Token efficiency and faster reasoning speed. It can switch to a low reasoning mode to significantly reduce Token consumption, and has been specially optimized for compatibility with Coding and Agent frameworks. Supports reasoning_effort field (low / high).

API Endpoints

Chat Completion

POST /v1/chat/completions OpenAI protocol compatible, supports streaming and tool calling.

Messages

POST /v1/messages Anthropic protocol compatible, can directly reuse Anthropic SDK.

Pricing

View full pricing details →

Quick Start

- curl

- Python (OpenAI SDK)

Related Reading

Reasoning Model Development Guide

Understand recommended usage of reasoning models in complex tasks, tool calling, and long contexts.

Step Router V1 Intelligent Routing

step-router-v1 routes to step-3.5-flash for execution when determining the task is high-frequency or deterministic.

This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.