证据: 基准测试结果由厂商自行报告
Step 3.7 Flash 198B-A11B — 显存、速度与本地部署
作者： Jakub Rusinowski · 最后更新：
196B 参数的 MoE 语言模型，加上 1.8B 视觉编码器（合计 198B），每个 token 约激活 11B：288 个路由专家，激活 8 个，另有共享专家；45 层中 12 层为全注意力、33 层为 512 token 滑动窗口；上下文 256K，三档推理强度。StepFun 发布的官方 GGUF 从 75.8 GB（IQ3_XXS）到 394 GB（BF16）；没有官方 Q4_K_M，因此此处的主要数字对应 Q4_K_S 文件。
Step 3.7 Flash 198B-A11B 在 Q4_K_S 下约需 112 GB 显存——量化权重加框架开销，不含 KV 缓存。在 Apple Silicon 上，这部分来自统一内存。
按量化级别的显存与速度
计算基准：NVIDIA RTX 4090 (24 GB)。包含 8K 上下文的 KV 缓存，因此数值高于上方的主数字。
| 量化 | 显存 | 显存 | 速度（估算） | 适配 |
|---|---|---|---|---|
| IQ3_XXS 3.09 bpw | 77.1 GB | — | 放不下 | |
| Q3_K_M 3.41 bpw | 95.1 GB | — | 放不下 | |
| Q3_K_L 4.18 bpw | 103.8 GB | — | 放不下 | |
| IQ4_XS 4.29 bpw | 106.3 GB | — | 放不下 | |
| Q4_K_S 4.55 bpw | 112.8 GB | — | 放不下 | |
| Q8_0 8.50 bpw | 210.8 GB | — | 放不下 | |
| BF16 16.00 bpw | 395.4 GB | — | 放不下 |
黑色标记 = NVIDIA RTX 4090 (24 GB) 上的可用显存。 估算来自内存带宽屋顶线模型，详见 方法说明页. Step 3.7 Flash 198B-A11B 显存计算器 →
运行 Step 3.7 Flash 198B-A11B
目录中能运行 Step 3.7 Flash 198B-A11B 的最便宜 GPU 是 NVIDIA DGX Spark (128 GB).
作为亚马逊联盟成员，我们从符合条件的购买中获得收入。云 GPU 链接为推荐链接——我们可能获得佣金，您无需额外付费。
如何运行 Step 3.7 Flash 198B-A11B
No Ollama tag. Use StepFun's official GGUFs with llama.cpp (mainline supports the Step 3.5 architecture the model uses; the card's build steps use StepFun's own fork), or vLLM / SGLang.
规格
Verified — Checked against the primary source — the model card or the vendor spec page — and corroborated by a second independent source.
- 参数量
- 198 Billion (11B active)
- 上下文窗口
- 262,144
- 架构
- Mixture-of-Experts (288 routed + 1 shared, 8 active), GQA with 3:1 sliding-window to full attention
- 提供商
- StepFun
- 许可证
- Apache-2.0
- Practical deployment
- Workstation (large unified memory or pro GPU)
- 规格量化
- Q4_K_S
- Vision projector
- +4 GB when you send images (not included above)
- 系统内存
- 128 GB
- 记录更新于
- 2026-11-03
Commercial use permitted. No usage restrictions beyond attribution.
限制与注意事项
- No official Q4_K_M build: the published 4-bit files are IQ4_XS (105.0 GB) and Q4_K_S (111.5 GB).
- StepFun recommends 128 GB of unified memory; at the default Apple Silicon GPU limit only the 3-bit files leave room for a cache.
- Benchmarks on the card are vendor-reported.
质量与使用场景
评分由模型作者或独立评测方发布——衡量质量而非吞吐量，并非我们实测。
StepFun Step 的其他尺寸
Step 3.7 Flash 198B-A11B — 常见问题
How much VRAM does Step 3.7 Flash 198B-A11B need?
About 112 GB at Q4_K_S — quantized weights plus framework overhead, before any KV cache. The cache grows with context length and is added on top; the table above folds it in. Apple Silicon counts unified memory toward the same figure.
Does Step 3.7 Flash 198B-A11B run on an RTX 4090 (24 GB)?
No. Step 3.7 Flash 198B-A11B needs about 112 GB at Q4_K_S, more than a single RTX 4090's 24 GB. It needs a larger card, several GPUs, or Apple Silicon with enough unified memory — or it runs with part of the weights offloaded to system RAM, which is much slower.
How do I run Step 3.7 Flash 198B-A11B locally?
No Ollama tag. Use StepFun's official GGUFs with llama.cpp (mainline supports the Step 3.5 architecture the model uses; the card's build steps use StepFun's own fork), or vLLM / SGLang. Running the published tag would send your prompts to a hosted GPU rather than your own machine.
What other sizes does StepFun Step come in?
Step 3.7 Flash 198B-A11B (112 GB), Step 5 Preview 600B-A27B (366 GB). Every size shares the family's training and licence; the larger ones score higher and need proportionally more memory.