Step 3.7 Flash – 198B-A11B MoE vision-language model (huggingface.co)

Step 3.7 Flash has been released as a cutting-edge 198 billion-parameter sparse Mixture-of-Experts (MoE) vision-language model, featuring a 196B language backbone paired with a 1.8B vision encoder. This model is optimized for high-frequency production tasks, activating around 11B parameters per token, and offers an impressive throughput of 400 tokens per second. With a context window of 256k and adjustable reasoning levels, it empowers developers to tailor performance to specific application needs, such as parsing complex financial reports and managing concurrent workflows. Significantly, Step 3.7 Flash excels in visual intelligence, achieving top scores on benchmarks like SimpleVQA and ClawEval-1.1, indicating superior capabilities in visual grounding and reliable execution in multi-step orchestration. Its ability to process dense visual interfaces, verify information, and generate structured code positions it as a powerful tool for tasks requiring nuanced understanding and precise interactions with external APIs. As a versatile solution adaptable across various environments, including cloud and local setups, Step 3.7 Flash is poised to enhance agentic workflows in fields ranging from software engineering to data analysis, ensuring high integrity and performance in autonomous agent applications.