On May 29, StepFun – the startup behind cutting-edge foundational AI models – launched and open-sourced their brand new Step 3.7 Flash. According to StepFun’s official WeChat post, this Flash model is tailor-made for production-level AI agents. Here’s the tech breakdown: it uses a sparse MoE architecture, with a total of 196B + 1.8B (ViT) parameters, but only 11B are activated. That means it can generate text at up to 400 tokens per second – pretty darn fast.
