Step 3.7 Flash is a 198B-parameter sparse Mixture-of-Experts (MoE) vision-language model engineered for high-frequency production workloads. It targets developers building agentic workflows, offering a balance of speed, cost, and cognitive depth with up to 400 tokens/sec throughput and a 256k context window.

How It Works

This model integrates a 196B-parameter language backbone with a 1.8B-parameter vision encoder, activating approximately 11B parameters per token. It supports selectable reasoning levels (low, medium, high) to tune performance against cost and depth requirements, enabling intensive tasks like parsing large documents, executing complex search loops, and managing concurrent coding agents.

Quick Start & Requirements

• Primary Install/Run: Available via StepFun API (platform.stepfun.ai, platform.stepfun.com), OpenRouter, NVIDIA NIM. Local deployment options include vLLM (Docker: vllm/vllm-openai:stepfun37), SGLang (Docker: lmsysorg/sglang:dev-step-3.7-flash), Hugging Face Transformers (requires transformers 5.0+), and llama.cpp.
• Prerequisites: High-memory devices (e.g., 128GB unified memory) are recommended for local deployment. NVFP4 variant requires NVIDIA GPUs. Specific build requirements exist for llama.cpp (Metal, CUDA, Vulkan).
• Resource Footprint: GGUF weights for llama.cpp are substantial (~102-111.5 GB LM + 3.97 GB projector), requiring minimum 120 GB unified memory/VRAM.
• Links:
  • Model Pages: BF16, FP8, NVFP4, GGUF.
  • StepFun Platform: platform.stepfun.ai.

Highlighted Details

• Achieves top-tier visual intelligence, ranking first on SimpleVQA (Search) and near-parity on V* (Python).
• Demonstrates high workflow integrity, leading the ClawEval-1.1 benchmark for agentic reliability.
• Offers an NVFP4-quantized variant optimized for NVIDIA GPUs, featuring MTP draft layers and supporting vLLM speculative decoding for improved throughput on long-reasoning tasks.
• Supports a 256k context window, enabling comprehensive analysis of large documents or complex multi-step workflows.

Maintenance & Community

• Active community engagement via Discord for brainstorming and feedback. Issues can be reported on GitHub/HuggingFace. Partnerships include DeepInfra, Fireworks AI, and Modal.

Licensing & Compatibility

• Licensed under the Apache 2.0 License, permissive for commercial use and closed-source linking.

Limitations & Caveats

• While strong in code engineering, benchmarks like Terminal-Bench 2.1 and GDPVal-AA indicate areas for future optimization compared to peak performers.