Skip to main content
Back

StepFun's Step-3.7-Flash Matches GPT-5.5 Vision at One-Ninth the Cost

StepFun's Step-3.7-Flash-GGUF brings a 198B multimodal MoE model to local hardware, with agentic tool-calling that rivals GPT-5.5 at a fraction of the cost

Model, Full Access
  • What: Step-3.7-Flash-GGUF is a fully quantized, locally runnable 198B multimodal MoE model from StepFun, under Apache 2.0.
  • Architecture: 198B total parameters, only ~11B active per token via sparse MoE routing; includes a 1.8B vision encoder and 256K context window.
  • Benchmarks: Scores 56.3 on SWE-Bench Pro and 67.1 on ClawEval-1.1 (first place), beating DeepSeek V4 Flash and Gemini 3.5 Flash on agentic tasks.
  • Local hardware: Runs on 128GB unified memory devices (Mac Studio M4 Max, DGX Spark, AMD Ryzen AI Max+ 395); Q4_K_S quant is 112GB.
  • Pricing: Cloud API at $0.20/M input and $1.15/M output tokens; $0.04/M on cache hits, making it cheap for high-frequency agent loops.
  • Advisor Mode: Achieves 97% of Claude Opus 4.6 coding performance at ~1/9th the cost ($0.19 vs $1.76 per agentic task).

Step-3.7-Flash-GGUF is StepFun's latest open-weight release: a fully quantized, locally runnable version of their 198B-parameter multimodal model, purpose-built for agentic workflows. It is available now on Hugging Face under an Apache 2.0 license, free for commercial use, and already pulling over 57,000 downloads.

A 198B model that thinks like an 11B one

Step-3.7-Flash is a 198 billion parameter sparse Mixture-of-Experts vision-language model that activates only 11 billion parameters per token, meaning it delivers frontier-class reasoning at a fraction of the compute cost. In a standard dense model, every single parameter fires on every token. MoE (Mixture-of-Experts) architectures instead maintain a large pool of specialized sub-networks and route each token through only a small subset. Only a subset of "expert" sub-networks fires per forward pass, keeping inference compute closer to an 11B dense model while maintaining a 198B total parameter budget.

Under the hood, it is a sparse Mixture-of-Experts transformer with 45 language layers, 42 of them routed MoE with 288 experts each and top-8 activation per token. It pairs a 196B-parameter language backbone with a 1.8B-parameter vision encoder (ViT) for native image understanding. The vision encoder ships as a separate 4GB mmproj file that you load alongside any of the language quantizations.

What it can actually do

This is not a general-purpose chat model. StepFun built it explicitly for agentic loops: multi-step reasoning, tool calling, code execution, and visual understanding. The benchmark story backs that up.