StepFun's Step-3.7-Flash Runs a 198B Model Like an 11B One
StepFun open-sources a 198B MoE vision-language model built for agentic coding, search, and multimodal workflows — with a Blackwell-native FP4 variant for maximum throughput
Step-3.7-Flash is StepFun's new open-weight flagship: a 198-billion-parameter sparse Mixture-of-Experts (MoE) vision-language model released under Apache 2.0 and purpose-built for agentic workloads. The model ships in four precision variants on Hugging Face , BF16, FP8, GGUF, and the NVFP4 checkpoint that targets NVIDIA Blackwell GPUs specifically , and has already crossed 220,000 downloads.
A 198B model that runs like an 11B one
Step 3.7 Flash is a 198B-parameter sparse MoE model that pairs a 196B-parameter language backbone with a 1.8B-parameter vision encoder for native image understanding. The MoE trick is what makes the scale practical: the model activates approximately 11B parameters per token during inference. In MoE architectures, only a subset of "expert" sub-networks fires per forward pass , not the full network , keeping inference compute closer to an 11B dense model while maintaining a 198B total parameter budget.
Engineered for high-frequency production workloads, it delivers throughput of up to 400 tokens per second and supports a 256k context window. It also offers three selectable reasoning levels , low, medium, and high , so developers can balance speed, cost, and cognitive depth. The model requires transformers 5.0 or later for deployment.
The NVFP4 variant: Blackwell-native 4-bit inference
The NVFP4 checkpoint is the most hardware-specific of the four variants. NVFP4 (NVIDIA FP4) is a 4-bit floating-point quantization format that is natively accelerated on Blackwell-class GPUs , meaning the GPU's tensor cores handle the 4-bit operations directly without dequantizing weights back to 16-bit first. NVFP4 avoids the dequantization bottleneck because Blackwell Tensor Cores are designed to handle NVFP4 operations directly, keeping tensors in their compact 4-bit format throughout inference, which means computations run much faster.