Run Step 3.7 Flash on NVIDIA GPUs with Enterprise-Ready Multimodal AI | NVIDIA Technical Blog - NVIDIA Developer

NVIDIA's recent technical blog highlights the launch of Step 3.7 Flash by StepFun, a multimodal AI model tailored for enterprise and production environments. This 198B-parameter Mixture-of-Experts vis... Read original →

Step 3.7 Flash by StepFun, powered by NVIDIA, signals the enterprise-grade arrival of multimodal AI. Technically, its 198B MoE design—activating only 11B params per pass—forces a re-architecting of inference stacks, demanding tighter integration between compilers (TensorRT-LLM, vLLM), memory schedulers, and HBM/NVLink subsystems. From a compliance standpoint, on-prem deployment via DGX Station sidesteps cross-border data restrictions, making it attractive for regulated sectors amid tightening U.S.-China AI controls. Competitively, this move pressures Google and Meta to accelerate enterprise-ready multimodal inference tooling or risk losing ground in vision-language workflows. Over the next 12–24 months, MoE will become the de facto architecture—but only ecosystems that unify training, fine-tuning, and optimized inference will dominate. NVIDIA’s NeMo + NIM + Blackwell stack currently stands alone. AI firms in Taiwan, China and Hong Kong, China face rising inference costs and delayed rollouts if A/H100 GPU access remains constrained.