Latest Content
Step 3.5 Flash icon

Step 3.5 Flash

NVIDIA
Step 3.5 Flash is a sparse Mixture-of-Experts (MoE) language model built for fast reasoning, coding, and agentic workflows. It has 196B total parameters with about 11B activated per token, using a 45-layer transformer with 4,096 hidden size and 64 attention heads. The architecture includes 288 experts per layer with top-8 routing and a shared expert, along with a hybrid attention pattern combining sliding window and full attention in a 3:1 ratio. It supports a 256K context window and integrates Multi-Token Prediction for high-throughput generation.
TypeMoE LLM
CapabilitiesText Generation, Instruction Following, Reasoning, Mathematical Reasoning+5 more
Release Date02 February, 2026
Links
Paper|Blog|HF Model Card
LicenseApache 2.0

Inference Instructions

Deploy and run this model on NVIDIA B200 GPUs using the command below. Copy the command to get started with inference.

CONSOLE
docker run -it --rm 
 --runtime=nvidia 
 --gpus all 
 --ipc=host 
 --shm-size=128g 
 -p 8000:8000 
 -v ~/.cache/huggingface:/root/.cache/huggingface 
 -e HF_TOKEN='YOUR_HF_TOKEN' 
 vllm/vllm-openai:v0.18.0 
 stepfun-ai/Step-3.5-Flash 
  --tensor-parallel-size 8 
 --max-model-len auto 
  --max-num-batched-tokens 65536 
 --disable-cascade-attn 
 --reasoning-parser step3p5 
 --tool-call-parser step3p5 
 --enable-auto-tool-choice 
 --gpu-memory-utilization 0.95 
 --max-num-seqs 1024 
 --trust-remote-code

Model Benchmarks

Each model was tested with a fixed input size and total token volume while increasing concurrency to measure serving performance under load.

ITL vs Concurrency

Time to First Token

Throughput Scaling

Total Tokens/sec vs Avg TTFT

Vultr Cloud GPU

NVIDIA HGX B200

Deploy NVIDIA B200 on Vultr Cloud GPU

Deploy NVIDIA B200 instancesContact Sales