Step 3.5 Flash
Step 3.5 Flash is an open-source, sparse Mixture-of-Experts (MoE) foundation model with 196B total parameters but only 11B active per token, designed for frontier reasoning and agentic tasks. It solves the problem of balancing high intelligence with fast, efficient inference, making it suitable for developers building AI agents, coding assistants, and real-time applications. It offers a 256K context window and can be deployed locally or accessed via cloud APIs.
✨ Key features
- Deep reasoning at speed with 100–300 tok/s throughput
- Purpose-built for coding and agentic tasks with scalable RL
- Efficient 256K context window using 3:1 sliding window attention
- Accessible local deployment on high-end consumer hardware
- Multi-Token Prediction (MTP-3) for faster inference
- Fine-grained MoE routing with 288 experts per layer
🎯 Use cases
- Building AI agents that require complex, multi-step reasoning
- Developing coding assistants that handle long codebases
- Deploying private, on-premise AI solutions with local inference
- Creating real-time interactive applications with low latency
- Running deep research and browsing tasks with context management
📦 Installation
🧰 Requirements: Requires Python 3.8+ and an OpenAI SDK for API access; for local deployment, needs vLLM, SGLang, Transformers, or llama.cpp with sufficient VRAM (e.g., 120GB for GGUF int4).
pip install --upgrade "openai>=1.0"
For local deployment with vLLM:
docker pull vllm/vllm-openai:nightlypip install -U vllm --pre \
--index-url https://pypi.org/simple \
--extra-index-url https://wheels.vllm.ai/nightlyFor SGLang:
docker pull lmsysorg/sglang:dev-pr-18084pip install "sglang[all] @ git+https://github.com/sgl-project/sglang.git"🚀 Usage
from openai import OpenAI
# Configuration for StepFun (International)
base_url = "https://api.stepfun.ai/v1"
api_key = "your-stepfun-api-key"
model = "step-3.5-flash"
client = OpenAI(api_key=api_key, base_url=base_url)
completion = client.chat.completions.create(
model=model,
messages=[
{"role": "system", "content": "You are a helpful assistant."},
{"role": "user", "content": "Hello! Introduce yourself."}
]
)
print(completion.choices[0].message.content)
⚠️ Good to know
Full MTP3 support is not yet available in vLLM; llama.cpp builds must include specific PRs for tool calling; China platform requires +86 phone verification.
❓ FAQ
What is the model's parameter count and active parameters?
It has 196.81B total parameters (196B backbone + 0.81B head) but only activates ~11B per token due to sparse MoE.
How can I access Step 3.5 Flash via API?
You can use OpenRouter (free and standard tiers) or StepFun's official API endpoints for International and China regions.
What are the hardware requirements for local deployment?
For llama.cpp GGUF int4, you need at least 120GB VRAM (e.g., Mac Studio, DGX Spark) and 128GB unified memory recommended.
Does it support tool calling and reasoning?
Yes, it supports tool calling and reasoning parsers in vLLM and SGLang, and is designed for agentic tasks.
📊 Repository
🤖 Overview, features, install steps and FAQ were generated from the project's README on Sep 4, 2026. Always check the original source before running commands.