Step-3.7-Flash-NVFP4
Step-3.7-Flash-NVFP4 is a 198B-parameter sparse mixture-of-experts vision-language model from StepFun that activates ~11B parameters per token. It processes images and text natively, supports a 256k context window, and delivers up to 400 tokens/sec throughput. The NVFP4 variant uses aggressive 4-bit quantization for lower memory footprint. Available via API (StepFun, OpenRouter, NVIDIA NIM) and deployable locally via vLLM, SGLang, or Transformers. Apache 2.0 licensed and not gated.
Key facts
Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.
| Field | Value |
|---|---|
| Developer | stepfun-ai |
| Parameters | 103.8B |
| Context window | Unknown |
| License | apache-2.0 — OSI-approved |
| Modality / task | image-text-to-text |
| Gated on HuggingFace | No |
| Downloads | 157.6k |
| Likes | 58 |
| Last updated | 2026-06-01 |
| Source | stepfun-ai/Step-3.7-Flash-NVFP4 |
What Step-3.7-Flash-NVFP4 is
A 198B-param sparse MoE vision-language model with 1.8B-param vision encoder and 196B-param language backbone. The NVFP4 variant applies NVIDIA ModelOpt 4-bit quantization. Supports 256k context, three reasoning levels (low/medium/high), and tool orchestration. Benchmarks show strong performance on SimpleVQA (79.2), ClawEval-1.1 (67.1), and SWE-Bench PRO (56.3). Requires FP8 KV cache alignment and tensor parallelism for inference. Supports speculative decoding (MTP/EAGLE) and expert parallelism.
Run Step-3.7-Flash-NVFP4 locally
Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.
from transformers import pipelinepipe = pipeline("text-generation", model="stepfun-ai/Step-3.7-Flash-NVFP4")out = pipe("Explain retrieval-augmented generation in one sentence.", max_new_tokens=128)print(out[0]["generated_text"])Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.
How you'd run it
A typical self-hosted path — open weights, an inference server, your application.
DEV.co builds each layer — from GPU infrastructure to the application.
Best use cases
Running & fine-tuning it
NVFP4 variant: 4-bit quantized model; exact VRAM not specified in card but 198B params at FP4 ~≤100 GB total (estimate requires verification). Requires FP8 KV-cache alignment per vLLM docs. Recommended: tensor-parallel-size ≥4 (documented example uses TP=4 with gpu-memory-utilization 0.9 on modern data center GPUs). Local deployment requires ≥128GB unified memory for non-cloud scenarios. BF16 variant would require ~396 GB (estimate); FP8 variant ~198 GB. Speculative decoding (MTP/EAGLE) recommended for throughput optimization.
Card does not describe LoRA, QLoRA, or supervised fine-tuning support. Mentions 'NVIDIA Megatron ecosystem, including Megatron Core and Megatron Bridge' for model development workflows, but fine-tuning feasibility, adapter method compatibility, and instruction-tuning guidance are not documented. Requires vendor or community guidance. Gating on custom code and trust-remote-code flags suggest model may use custom layers (MoE) that complicate adaptation.
When to avoid it — and what to weigh
- Strict Real-Time Latency Requirements (<50ms) — While 400 tok/sec is high-throughput, inference latency depends on tensor-parallel deployment scale and KV-cache configuration. Requires careful benchmarking for hard real-time SLAs; not suitable for interactive sub-50ms response demands without speculative decoding tuning.
- Edge Deployment on Consumer Hardware — NVFP4 variant still requires significant GPU memory. Local deployment documented only for high-memory devices (NVIDIA DGX Station, AMD Ryzen AI Max+, Mac Studio with 128GB unified memory minimum). Not practical for edge devices or single-GPU workstations.
- Low-Context or Simple Q&A Workloads — 128B+ active parameters and MoE overhead make this overprovisioned for simple chatbots or factual lookup tasks. Cost-per-token (especially $1.15/M output) and complexity justify use only for agentic, reasoning-heavy, or visual workflows.
- Mission-Critical Deployments Without Extensive In-House Benchmarking — Benchmarks provided by StepFun; independent third-party validation not cited. Terminal-Bench 2.1 (59.5) and GDPVal-AA (45.8) show gaps versus cohort leaders. Requires organization to validate performance on own tasks before production commitments.
License & commercial use
Apache 2.0 license. Permissive OSI-approved open-source license permits commercial use, modification, and redistribution with attribution and license notice. No copyleft clause. Model weights and code are not gated.
Apache 2.0 is a permissive OSI license that explicitly permits commercial use. No additional proprietary agreement, commercial seat license, or usage restrictions are mentioned in the card. Commercial deployment via self-hosted (vLLM, SGLang, Transformers) or via managed API (StepFun, OpenRouter, NVIDIA NIM) is permitted. Pricing is published ($0.20/M input cache miss, $0.04/M cache hit, $1.15/M output) for API use. Verify with StepFun directly if deploying at scale for terms on data residency, SLAs, and support.
DEV.co evaluation signals
Editorial assessment — not user reviews. Directional, with an explicit confidence level.
| Signal | Assessment |
|---|---|
| Maintenance | Unknown |
| Documentation | Adequate |
| License clarity | Clear |
| Deployment complexity | Moderate |
| DEV.co fit | Good |
| Assessment confidence | Medium |
Card does not discuss adversarial robustness, prompt injection mitigations, data retention, or model extraction safeguards. High score on ClawEval-1.1 adversarial benchmark (67.1) suggests some robustness to policy-violation traps, but this is a model-level benchmark, not infrastructure security. Tool orchestration features (auto-tool-choice, tool-call-parser) introduce dependency on external API correctness and sandboxing. Local deployment requires trust-remote-code flag in vLLM/SGLang, indicating custom layers; audit required before use. API deployments (StepFun, OpenRouter, NVIDIA NIM) shift security responsibility to vendor. No mention of SOC 2, data encryption, audit logging, or DLP controls.
Alternatives to consider
Claude 3.5 Sonnet (Anthropic)
Multimodal generalist with 200k context; lower latency; strong on code and reasoning. Proprietary; managed API only; higher per-token cost. Choose if you need industry-leading safety/alignment and can accept vendor lock-in.
GPT-4o (OpenAI)
Multimodal, strong vision+code; 128k context; fastest inference among frontier models. Proprietary; managed API only; highest per-token cost. Choose for absolute peak performance on visual reasoning and broad use cases.
Mixtral 8x22B (Mistral, open-source)
Open-source MoE sparse model (141B params, ~39B active); Apache 2.0 licensed; deployable on-prem. No vision; lower benchmark scores than Step-3.7-Flash. Choose if vision is not required and full data sovereignty is critical.
Ship Step-3.7-Flash-NVFP4 with senior software developers
Start with a free API tier on StepFun Platform or test locally using the vLLM Docker image. Benchmark the NVFP4 quantized variant against your own tasks (visual understanding, code, tool orchestration) to validate performance before production deployment.
Talk to DEV.coRelated open-source tools
Surfaced by semantic similarity across the DEV.co open-source index.
Related on DEV.co
Explore the category and the services that help you build with it.
Step-3.7-Flash-NVFP4 FAQ
Can I use Step-3.7-Flash-NVFP4 for commercial products without additional licensing?
What GPU hardware is required to run the NVFP4 variant locally?
How does the NVFP4 quantization affect accuracy vs. the full-precision BF16 model?
Can I fine-tune or adapt Step-3.7-Flash-NVFP4 for domain-specific tasks?
Software development & web development with DEV.co
Adopting Step-3.7-Flash-NVFP4 is usually one piece of a larger software development effort. As a software development agency, DEV.co provides software development services and web development expertise — pairing senior software developers and web developers with your team to design, build, and operate open-source llms software in production.
Evaluate Step-3.7-Flash for Your Agentic Workflow
Start with a free API tier on StepFun Platform or test locally using the vLLM Docker image. Benchmark the NVFP4 quantized variant against your own tasks (visual understanding, code, tool orchestration) to validate performance before production deployment.