DEV.co
Open-Source LLM · stepfun-ai

Step-3.7-Flash

Step-3.7-Flash is a 198B-parameter sparse Mixture-of-Experts vision-language model that activates ~11B parameters per token, delivering up to 400 tokens/second throughput. It supports 256k context window, native image understanding, and three selectable reasoning levels (low/medium/high). Designed for production agentic workflows combining perception, search, and reasoning. Available via StepFun's API, OpenRouter, NVIDIA NIM, and self-hosted deployment on modern infrastructure.

Source: HuggingFace — huggingface.co/stepfun-ai/Step-3.7-Flash
201.4B
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
146.5k
Downloads (30d)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
Developerstepfun-ai
Parameters201.4B
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / taskimage-text-to-text
Gated on HuggingFaceNo
Downloads146.5k
Likes406
Last updated2026-06-03
Sourcestepfun-ai/Step-3.7-Flash

What Step-3.7-Flash is

Step-3.7-Flash is a 198B MoE vision-language model with a 196B language backbone and 1.8B vision encoder. It activates approximately 11B parameters per token, achieving ~400 tokens/second throughput. Supports 256k context window and three reasoning levels. Benchmark performance: ClawEval-1.1 (67.1 — agentic workflow reliability), SWE-Bench PRO (56.3 — code engineering), SimpleVQA (79.2 — visual grounding), V*-Python (95.3 — code generation from visual assets). Available via vLLM, SGLang, Hugging Face Transformers, llama.cpp, and NVIDIA NIM. Licensed under Apache 2.0.

Quickstart

Run Step-3.7-Flash locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="stepfun-ai/Step-3.7-Flash")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.

Deployment

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

High-frequency production agentic pipelines

Step-3.7-Flash is engineered for scale: 400 tokens/second throughput and 11B active parameters per token minimize latency and cost. Ideal for concurrent coding agents, multi-turn search orchestration, and autonomous workflows that require strict instruction adherence and tool integration reliability (ClawEval-1.1: 67.1).

Multimodal document parsing and verification

Parse dense financial reports, invoices, and data-rich UI assets in a single pass. Strong visual grounding (SimpleVQA: 79.2) and independent lookup capability enable factually verified conclusions without repeated round-trips. 256k context window accommodates large documents.

Software engineering and code generation

SWE-Bench PRO score of 56.3 demonstrates ability to trace multi-file repositories, isolate bugs from issue reports, and generate patches that pass unit tests. Terminal-Bench (59.5) and GDPVal-AA (45.8) provide baseline system interaction capability for structured professional deliverables.

Running & fine-tuning it

Step-3.7-Flash is a 198B-parameter sparse MoE model. Self-hosted inference requires: (1) For FP8 quantization: estimate 400–600GB VRAM across tensor-parallel GPU setup (recommend 8× H100 or equivalent); (2) For BF16: estimate 500–1000GB VRAM; (3) For NVFP4 (quantized): estimate 250–400GB VRAM with FP8 KV cache. Local workstation deployment (vLLM/SGLang) requires ≥128GB unified memory (DGX Station, Mac Studio, Ryzen AI Max+). API-based inference eliminates hardware burden but incurs per-token costs.

Fine-tuning capability is not explicitly stated in the model card. Model is available via NVIDIA NeMo ecosystem (AutoModel, Megatron Core, Megatron Bridge) for 'model development & customization workflows,' suggesting some fine-tuning support is plausible. LoRA/QLoRA feasibility Unknown — requires review of NeMo documentation or StepFun's official guidance. Custom code flagged in model tags; verify licensing and technical constraints before attempting adaptation.

When to avoid it — and what to weigh

  • Absolute peak-performance code generation required — While competitive (SWE-Bench PRO: 56.3), Step-3.7-Flash is positioned as a 'definitive second-place' model. Terminal-Bench 2.1 (59.5) and GDPVal-AA (45.8) lag the absolute cohort leaders. If top-tier code engineering is non-negotiable, verify against specific benchmarks for your task.
  • Fully on-premises deployment with minimal infrastructure — Local deployment requires high-memory devices: NVIDIA DGX Station, AMD Ryzen AI Max+ 395, or Mac Studio/MacBook Pro with ≥128GB unified memory. Sparse MoE and 198B parameters demand modern, well-equipped infrastructure. Estimate GPU VRAM requirements at 500–1000GB for inference depending on precision (FP8/BF16/FP4) — verify with vLLM/SGLang configs.
  • Commercial deployment without API terms review — Apache 2.0 license permits commercial use, but deployment through StepFun's API or NVIDIA NIM is subject to their respective terms and pricing ($0.20/M input, $1.15/M output). Self-hosted deployment avoids vendor terms but requires significant engineering and infrastructure investment.
  • Multi-language production use beyond stated support — Model card states support for Chinese, English, and 'many other languages,' but evaluation benchmarks are English-centric. Production use in non-English contexts requires independent validation.

License & commercial use

Step-3.7-Flash is licensed under Apache 2.0, an OSI-approved permissive open-source license. Apache 2.0 permits commercial use, modification, and distribution with minimal restrictions (includes patent protection grant and liability disclaimer).

Apache 2.0 license explicitly permits commercial use of the model weights. However: (1) API-based inference via StepFun, OpenRouter, NVIDIA NIM, DeepInfra, Fireworks AI, or Modal is subject to those platforms' commercial terms and pricing ($0.20/M input tokens cache miss, $0.04/M cache hit, $1.15/M output). Review vendor agreements before committing to production use. (2) Self-hosted deployment on your infrastructure carries no licensing restrictions under Apache 2.0 but requires significant operational overhead (GPUs, vLLM/SGLang setup, monitoring).

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationAdequate
License clarityClear
Deployment complexityHigh
DEV.co fitStrong
Assessment confidenceHigh
Security considerations

Model card contains no explicit security audit, red-teaming results, or vulnerability disclosures. Considerations: (1) Tool orchestration and API integration: ClawEval-1.1 score (67.1) indicates 'high resistance to adversarial traps,' but adversarial robustness is not absolute; test tool-call safety in your sandbox before production. (2) Custom code flagged in model tags; vLLM/SGLang deployment requires --trust-remote-code, which executes model-provided Python—audit before use in untrusted environments. (3) Vision encoder may inherit risks from image processing pipelines (adversarial images, encoded exploits); isolate image inputs if handling untrusted user uploads. (4) API-based deployment: data traverses StepFun infrastructure; review data residency and privacy policies if handling sensitive information. No formal security certification or compliance badges mentioned.

Alternatives to consider

Claude 3.5 Sonnet (Anthropic)

Closed-source, API-only multimodal model with strong code engineering and reasoning. Trade-off: proprietary, higher per-token costs, no self-hosting. Best if you prioritize maximum capability over operational control.

Llama 3.2 Vision (Meta, OSS)

Open-source multimodal model, smaller parameter count (~11B), lower inference cost. Trade-off: smaller context window, weaker agentic workflow scores, but fully transparent and self-hostable. Best for cost-sensitive use cases with modest scale.

GPT-4 Turbo with Vision (OpenAI)

Closed-source, API-only, mature production track record. Trade-off: proprietary, expensive, no self-hosting. Best if vendor stability and broad compatibility are paramount.

Software development agency

Ship Step-3.7-Flash with senior software developers

Assess your infrastructure needs: estimate GPU VRAM (400–1000GB depending on precision), review API pricing vs. self-hosted TCO, and validate agentic workflow benchmarks (ClawEval-1.1, SWE-Bench PRO) against your production use case. Start with the vLLM Docker image or test via StepFun's API before committing large-scale inference.

Talk to DEV.co

Related 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 FAQ

Can I use Step-3.7-Flash commercially without restrictions?
Yes, the model weights are Apache 2.0 licensed, permitting commercial use. However, if you deploy via StepFun's API, OpenRouter, NVIDIA NIM, or other vendors, you must comply with their commercial terms and pricing. Self-hosted deployment carries no licensing restrictions but requires significant infrastructure investment and operational expertise.
What GPU(s) do I need to self-host Step-3.7-Flash?
For self-hosted inference, estimate 400–600GB VRAM for FP8 quantization (recommend 8× H100 or equivalent). BF16 precision requires ~500–1000GB VRAM. NVFP4 quantization reduces to ~250–400GB VRAM with FP8 KV cache. Workstation deployment requires ≥128GB unified memory (DGX Station, Mac Studio, Ryzen AI Max+). Use vLLM or SGLang for orchestration. These estimates assume tensor parallelism and batch inference; verify with actual deployment.
How does Step-3.7-Flash compare to Claude or GPT-4 for code generation?
SWE-Bench PRO score of 56.3 places Step-3.7-Flash as competitive but not peak-tier. It is sufficient for tracing multi-file repos and generating patches that pass tests, but if absolute peak code engineering is critical, verify against specific task benchmarks. Claude 3.5 and GPT-4 Turbo likely exceed it on proprietary benchmarks, but closed-source evaluation claims lack independent verification.
Does Step-3.7-Flash support fine-tuning or LoRA?
Fine-tuning capability is not explicitly documented in the model card. NVIDIA NeMo ecosystem integration (Megatron Core, Megatron Bridge) suggests some customization pathway exists, but details are Unknown. Contact stepfun-ai or review NeMo documentation for formal support. LoRA/QLoRA feasibility requires independent verification.

Custom software development services

Adopting Step-3.7-Flash 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.

Ready to Deploy Step-3.7-Flash?

Assess your infrastructure needs: estimate GPU VRAM (400–1000GB depending on precision), review API pricing vs. self-hosted TCO, and validate agentic workflow benchmarks (ClawEval-1.1, SWE-Bench PRO) against your production use case. Start with the vLLM Docker image or test via StepFun's API before committing large-scale inference.