StepFun's Step-3.5-Flash is a 196.8B MoE at 256K context, Apache-2.0. Beyond retail GPUs; needs multi-GPU or datacenter-class hardware.
Step is the model family from StepFun (阶跃星辰), a Chinese AI lab that has become a community favorite for big open-weight releases. On this site the family currently has one headline member: Step-3.5-Flash, a sparse MoE model released in 2026 under the permissive Apache-2.0 license.
The honest headline up front: this is not a single-GPU hobbyist model. With 196.8B total parameters, even a Q4 quant needs datacenter-class or carefully assembled multi-GPU hardware. The good news is the MoE design — only about 11B parameters are active per token — so once you fit it in VRAM, decode speed feels closer to a 13B-class dense model than a 200B monster.
Size ladder
- Step-3.5-Flash (196.8B total, MoE with 11B active; 288 experts, 8 routed per token) — Q4 ≈ 111 GB VRAM at the default 8K context → beyond every 96GB-class card; you need a 192GB-class setup such as a single MI300X 192GB, B200, or Mac Studio M2 Ultra 192GB, or a multi-GPU rig. FP16 weights alone are ≈ 394 GB. See GPU fit & speed estimates
What makes the Step family special
- Sparse MoE efficiency. 288 experts with 8 active per token means quality in the 200B league while compute per token stays in the ~11B league. That ratio is exactly why "Flash" earns its name for serving.
- 256K context. One of the longest context windows in the open-weight world. Keep in mind the KV cache grows with context: at the full 256K, KV alone adds roughly 70 GB on top of the ~111 GB you need at 8K. Long-document workloads should budget 192GB or more.
- Apache-2.0 license. Free for commercial use, fine-tuning, and redistribution — a real differentiator versus restricted big-model licenses.
- GGUF-friendly. Community GGUF quants circulate widely, so llama.cpp-style local serving with CPU offload is practical if you accept slower tok/s.
Which one should you pick
Since the family currently means one model, the real question is whether your hardware tier can hold it:
- 12GB / 16GB / 24GB cards — not in the conversation. Even the smallest quants of a 196.8B model will not fit. Pick a smaller model from our best lists instead.
- 48GB (RTX 6000 Ada / A6000) — still no: even a Q2 quant needs ≈ 58 GB. Two 48GB cards (96 GB pooled) can hold Q2 with room for an 8K context.
- 96GB class (Mac Studio M3 Ultra 96GB) — Q2 fits (~58 GB); Q4 (~111 GB) does not. Q2 is usable for testing, but expect visible quality loss on hard prompts.
- 128–141GB class (MacBook M4 Max 128GB, H200 141GB) — Q4 fits at 8K context with some headroom, but 256K context will not. A sweet spot for single-machine Q4 serving at moderate context.
- 192GB+ (MI300X, B200, Mac Studio M2 Ultra 192GB, 2×96GB) — the comfortable answer. Q4 fits at 8K and leaves room for long-context KV; this is the tier we recommend for serious local or on-prem use.
- FP16 purists — weights alone are ≈ 394 GB, so you are shopping for a multi-GPU server node, not a workstation.
Links
1 models in the Step family, grouped by series. Q4 (GB) is weights-only; total VRAM adds KV cache and overhead — the fit check links compute it for a representative retail GPU at 8K context.