Skip to main content
Back

NVIDIA Wires StepFun's 198B Step 3.7 Flash Into Its Full Stack

NVIDIA rolls out Day 0 support for StepFun's Step 3.7 Flash, with GPU endpoints, NIM microservices, and NeMo fine-tuning recipes ready at launch.

  • NVIDIA enables Day 0 support for StepFun's Step 3.7 Flash across endpoints, NIM, and NeMo.
  • Step 3.7 Flash is a 198B MoE vision-language model with 11B active parameters and 256k context.
  • GPU-accelerated endpoints on build.nvidia.com let developers prototype before self-hosting.
  • NIM containers package the model with an OpenAI-compatible API for on-prem or hybrid deployment.
  • NeMo Automodel supports LoRA fine-tuning at 600 tokens/sec on Hopper, no checkpoint conversion needed.
  • NVFP4-quantized checkpoint available on Hugging Face reduces memory bandwidth requirements.

NVIDIA has switched on Day 0 support for StepFun's newest open multimodal model, Step 3.7 Flash, meaning the model can be prototyped, deployed, and fine-tuned across NVIDIA's developer stack the moment its weights hit Hugging Face. The rollout spans GPU-accelerated endpoints on build.nvidia.com, packaged NIM inference microservices, and ready-made NeMo fine-tuning recipes, removing most of the integration work that typically follows an open model release.

A vision-language model wired into NVIDIA's stack on arrival

Step 3.7 Flash itself is StepFun's latest vision-language model. It is a 198B-parameter Mixture-of-Experts model with approximately 11B activated parameters per forward pass, optimized for agentic workflows that combine perception, search, and multi-step reasoning, and it ships with native image and video input, three configurable reasoning levels, and a 256k context window.

What NVIDIA is announcing is not the model, but the fact that its infrastructure is ready for it on day one. Developers can pull StepFun's NVFP4-quantized checkpoint from Hugging Face for boosted inference thanks to reduced memory bandwidth and storage requirements, and run it through the open-source serving stacks that NVIDIA maintains kernels for.

Three ways to plug it in

The technical blog lays out three entry points, mapped to different stages of the development cycle:

  • Hosted endpoints for prototyping. build.nvidia.com provides GPU-accelerated endpoints for quick prototyping, and a demo notebook pairs Step 3.7 Flash with NVIDIA Nemotron Parse to showcase a document-intelligence pipeline that extracts structured data from PDFs, slide decks and scientific papers.
  • Open-source serving frameworks. The model can be deployed through SGLang, NVIDIA TensorRT-LLM, and vLLM, which tap kernels tuned for NVIDIA hardware.
  • NIM microservices for production. NVIDIA NIM enables production-ready deployment of Step 3.7 Flash as containerized inference microservices with standardized APIs, supporting on-premises, cloud, and hybrid environments, alongside Day 0 fine-tuning capabilities using the NVIDIA NeMo framework for domain-specific customization.

The NIM container exposes an OpenAI-compatible API, so existing client code typically only needs a base URL swap:

from openai import OpenAI
client = OpenAI(
    base_url="http://0.0.0.0:8000/v1",
    api_key="no-key-required",
)
completion = client.chat.completions.create(
    model="stepfun/step-3.7-flash",
    messages=[{"role": "user", "content": "Explain particle physics?"}],
    max_tokens=1024,
    stream=True,
)

Fine-tuning without checkpoint conversion

The NeMo side is where the Day 0 framing matters most. NVIDIA's NeMo Automodel library combines native PyTorch n-D parallelisms with optimized performance and supports fine-tuning directly from Hugging Face checkpoints without conversion. The Automodel recipe for Step 3.7 supports supervised fine-tuning and memory-efficient LoRA at 600 tokens per second on Hopper GPUs, and a Megatron-Bridge recipe is available for larger-scale training runs.

For model development and customization workflows, StepFun model support has landed in the NVIDIA NeMo ecosystem, including AutoModel, Megatron Core and Megatron Bridge, and Step 3.7 Flash is also available as an NVIDIA NIM inference microservice for on-prem, cloud, or hybrid deployment. That coverage from inference to training is what the Day 0 label actually buys teams.

Why this matters for multimodal pipelines

The reference workload NVIDIA highlights is document intelligence. The multi-step document intelligence pipeline extracts structured insights from large, complex documents with bounding boxes like financial reports, slide decks, and scientific papers, including PDFs, and organizes the output. That combination of perception plus reasoning is exactly the kind of agentic workload Step 3.7 Flash is targeted at, and it now has a turnkey path through NVIDIA's stack.

For teams building multimodal agents, the practical takeaway is that the choice between hosted, self-hosted, or fine-tuned deployments no longer has to be made sequentially. The same Hugging Face checkpoint can be tried on build.nvidia.com, dropped into a NIM container for staging, and pushed through a NeMo LoRA recipe for domain customization without rebuilding the pipeline at each step. For an open model release, that is what Day 0 enterprise support is supposed to look like.