Step 3.7 Flash
Step 3.7 Flash is a 198B-parameter sparse Mixture-of-Experts vision-language model with a 1.8B vision encoder, designed for high-frequency production workloads. It activates ~11B parameters per token, supports a 256k context window, and offers three reasoning levels. It is for developers building agentic workflows that combine perception, search, and reasoning.
✨ Key features
- Native image understanding with vision encoder
- Three selectable reasoning levels (low, medium, high)
- 256k context window
- NVFP4 quantization with MTP speculative decoding
- Supports vLLM, SGLang, Transformers, llama.cpp
- Available via API and local deployment
🎯 Use cases
- Parsing massive financial reports in one pass
- Running multi-step search loops with cross-source verification
- Operating concurrent coding agents in high-throughput pipelines
- Mapping UI wireframes and data charts to structured code
- Building autonomous agents that interact with external APIs
📦 Installation
🧰 Requirements: Requires Python, an API key from StepFun (or other providers), and for local deployment, high-memory hardware (e.g., 128GB unified memory) or NVIDIA GPUs with sufficient VRAM.
For API usage, install the OpenAI Python client:
pip install openai
For local deployment with vLLM, use the prebuilt Docker image:
docker pull vllm/vllm-openai:stepfun37
For SGLang, use Docker or pip:
docker pull lmsysorg/sglang:dev-step-3.7-flashpip install "sglang[all] @ git+https://github.com/sgl-project/sglang.git"For Transformers, ensure transformers 5.0+ is installed.
For llama.cpp, clone the StepFun fork and build:
git clone https://github.com/stepfun-ai/llama.cpp.git
cd llama.cpp
git checkout -b step3.7 origin/step3.7
🚀 Usage
Here is a minimal chat example using the StepFun API:
import os
from openai import OpenAI
client = OpenAI(
api_key=os.environ["STEP_API_KEY"],
base_url=os.environ["STEP_BASE_URL"],
)
completion = client.chat.completions.create(
model="step-3.7-flash",
messages=[
{"role": "system", "content": "You are an AI assistant provided by StepFun."},
{"role": "user", "content": "Introduce StepFun's artificial intelligence capabilities."}
],
)
print(completion)
Set environment variables first:
export STEP_API_KEY="sk-..."
export STEP_BASE_URL="https://api.stepfun.ai/v1" # or https://api.stepfun.com/v1 for China
⚠️ Good to know
The README notes that evaluations like Terminal-Bench 2.1 and GDPVal-AA show areas for future optimization compared to the peak of the cohort, and the NVFP4 benchmark characterizes short-prompt, decode-heavy reasoning rather than long-context prefill performance.
❓ FAQ
What is the model size and parameter count?
Step 3.7 Flash has 198B total parameters, with a 196B language backbone and a 1.8B vision encoder, but activates only ~11B parameters per token.
How can I access the model?
It is available via the StepFun Open Platform (platform.stepfun.ai for global, platform.stepfun.com for China), OpenRouter, and NVIDIA NIM. It can also be deployed locally using vLLM, SGLang, Transformers, or llama.cpp.
What are the pricing details?
Input tokens cost $0.20 per million on cache miss and $0.04 per million on cache hit; output tokens cost $1.15 per million.
What hardware is needed for local deployment?
For llama.cpp, you need at least 120GB unified memory or VRAM, with 128GB recommended. For vLLM, you need GPUs with sufficient VRAM; the examples use tensor-parallel-size 8 for BF16/FP8 and 4 for NVFP4.
📊 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.