Step 3.7 Flash

Step 3.7 Flash is a Mixture of Experts (MoE) large language model with 198B total parameters but only 11B active per token developed by StepFun. Released in June 2026, it supports text and vision inputs with a 256K context window, released under the Apache 2.0 license, allowing commercial use. Multimodal successor to Step 3.5 Flash, retaining the 196B language backbone and adding an approximately 1.8B-parameter vision encoder with low, medium, and high reasoning levels. StepFun reports SWE-bench Pro 56.3 and Terminal-Bench 2.1 59.5. Official GGUF releases are about 105-112 GB before the roughly 4 GB vision projector and runtime overhead; StepFun also provides an NVFP4 checkpoint.

To run Step 3.7 Flash locally, you need approximately 136.7 GB of VRAM at Q4_K_M quantization with 8k context. 20 of the GPUs we track can run it fully in VRAM, with a further 2 able to offload to system RAM. At Q4_K_M it requires 136.7 GB — more than any single consumer GPU. A multi-GPU server or 80 GB datacenter GPU is required. At Q8_K_M (237.4 GB), you get near-FP16 quality while still fitting on large server or multi-GPU setups. FP16 requires 445.2 GB, limiting it to datacenter-class hardware with 80 GB+ VRAM. As a MoE model, inference speed depends on active parameters (11B) rather than total size, so once it fits in VRAM it runs noticeably faster than dense models of the same parameter count.

Step 3.7 Flash is designed for reasoning and problem-solving. Thanks to its MoE architecture with only 11B active parameters, it delivers strong quality while keeping inference latency manageable for local use.