{
  "hf_id": "stepfun-ai/Step-3.7-Flash",
  "meta": {
    "title": "Step-3.7-Flash",
    "slug": "step-3.7-flash",
    "provider": "StepFun",
    "description": "Production-grade vision-language MoE (~198B total / 11B active parameters) combining a 196B sparse language backbone with a 1.8B perception encoder, hybrid SWA/Global attention, and 3-way Multi-Token Prediction",
    "date_added": "2026-05-29",
    "date_updated": "2026-05-30",
    "tasks": [
      "multimodal"
    ],
    "performance_headline": "Sparse MoE VLM with hybrid attention and 3-layer MTP speculative decoding",
    "related_recipes": [
      "stepfun-ai/Step-3.5-Flash"
    ],
    "hardware": {
      "h200": "verified",
      "b200": "verified",
      "dgx_station_gb300": "verified"
    }
  },
  "recommended_command": {
    "hardware": "h200",
    "strategy": "single_node_tp",
    "variant": "default",
    "node_count": 1,
    "deploy_type": "single_node",
    "env": {},
    "docker_image": "vllm/vllm-openai:stepfun37",
    "command": "vllm serve stepfun-ai/Step-3.7-Flash \\\n  --trust-remote-code \\\n  --enable-expert-parallel \\\n  --disable-cascade-attn \\\n  --tensor-parallel-size 8 \\\n  --tool-call-parser step3p5 \\\n  --enable-auto-tool-choice \\\n  --reasoning-parser step3p5",
    "argv": [
      "vllm",
      "serve",
      "stepfun-ai/Step-3.7-Flash",
      "--trust-remote-code",
      "--enable-expert-parallel",
      "--disable-cascade-attn",
      "--tensor-parallel-size",
      "8",
      "--tool-call-parser",
      "step3p5",
      "--enable-auto-tool-choice",
      "--reasoning-parser",
      "step3p5"
    ],
    "docker_command": "docker run --gpus all \\\n  --privileged --ipc=host -p 8000:8000 \\\n  -v ~/.cache/huggingface:/root/.cache/huggingface \\\n  vllm/vllm-openai:stepfun37 stepfun-ai/Step-3.7-Flash \\\n  --trust-remote-code \\\n  --enable-expert-parallel \\\n  --disable-cascade-attn \\\n  --tensor-parallel-size 8 \\\n  --tool-call-parser step3p5 \\\n  --enable-auto-tool-choice \\\n  --reasoning-parser step3p5",
    "docker_argv": [
      "docker",
      "run",
      "--gpus",
      "all",
      "--privileged",
      "--ipc=host",
      "-p",
      "8000:8000",
      "-v",
      "~/.cache/huggingface:/root/.cache/huggingface",
      "vllm/vllm-openai:stepfun37",
      "stepfun-ai/Step-3.7-Flash",
      "--trust-remote-code",
      "--enable-expert-parallel",
      "--disable-cascade-attn",
      "--tensor-parallel-size",
      "8",
      "--tool-call-parser",
      "step3p5",
      "--enable-auto-tool-choice",
      "--reasoning-parser",
      "step3p5"
    ],
    "strategy_spec": {
      "name": "single_node_tp",
      "deploy_type": "single_node",
      "display_name": "Tensor Parallel",
      "short_label": "TP",
      "orientation": "latency",
      "description": "Single-node tensor parallel. Splits the model across all local GPUs. TP size is set to the GPU count at deploy time. The simplest multi-GPU strategy — works for all model architectures.\n",
      "hardware_match": {
        "min_gpus": 1,
        "max_gpus": 8,
        "multi_node": false
      },
      "vllm_args": [],
      "parallel_flag": "--tensor-parallel-size"
    },
    "hardware_profile": {
      "brand": "NVIDIA",
      "generation": "hopper",
      "compute_arch": "sm90",
      "display_name": "H200",
      "description": "NVIDIA H200 SXM 141 GB HBM3e · 8-GPU HGX node",
      "gpu_count": 8,
      "vram_gb": 1128,
      "multi_node": false
    },
    "alternatives": {
      "single_node_tep": "/stepfun-ai/Step-3.7-Flash/strategies/single_node_tep.json",
      "single_node_dep": "/stepfun-ai/Step-3.7-Flash/strategies/single_node_dep.json",
      "multi_node_tp": "/stepfun-ai/Step-3.7-Flash/strategies/multi_node_tp.json",
      "multi_node_tep": "/stepfun-ai/Step-3.7-Flash/strategies/multi_node_tep.json",
      "multi_node_dep": "/stepfun-ai/Step-3.7-Flash/strategies/multi_node_dep.json",
      "pd_cluster": "/stepfun-ai/Step-3.7-Flash/strategies/pd_cluster.json",
      "kv_store_centralized_mooncake": "/stepfun-ai/Step-3.7-Flash/strategies/kv_store_centralized_mooncake.json",
      "kv_store_distributed_mooncake": "/stepfun-ai/Step-3.7-Flash/strategies/kv_store_distributed_mooncake.json"
    },
    "by_hardware": {
      "h200": "/stepfun-ai/Step-3.7-Flash/hw/h200.json",
      "h100": "/stepfun-ai/Step-3.7-Flash/hw/h100.json",
      "b200": "/stepfun-ai/Step-3.7-Flash/hw/b200.json",
      "gb200": "/stepfun-ai/Step-3.7-Flash/hw/gb200.json",
      "b300": "/stepfun-ai/Step-3.7-Flash/hw/b300.json",
      "gb300": "/stepfun-ai/Step-3.7-Flash/hw/gb300.json",
      "dgx_station_gb300": "/stepfun-ai/Step-3.7-Flash/hw/dgx_station_gb300.json",
      "mi300x": "/stepfun-ai/Step-3.7-Flash/hw/mi300x.json",
      "mi325x": "/stepfun-ai/Step-3.7-Flash/hw/mi325x.json",
      "mi350x": "/stepfun-ai/Step-3.7-Flash/hw/mi350x.json",
      "mi355x": "/stepfun-ai/Step-3.7-Flash/hw/mi355x.json"
    }
  },
  "model": {
    "model_id": "stepfun-ai/Step-3.7-Flash",
    "min_vllm_version": "0.23.0",
    "docker_image": "vllm/vllm-openai:stepfun37",
    "architecture": "moe",
    "parameter_count": "198B",
    "active_parameters": "11B",
    "context_length": 262144,
    "base_args": [
      "--trust-remote-code",
      "--enable-expert-parallel",
      "--disable-cascade-attn"
    ],
    "base_env": {},
    "install": {
      "pip": {
        "command": "uv venv\nsource .venv/bin/activate\nuv pip install -U vllm --torch-backend auto"
      },
      "docker": {
        "command": "docker pull vllm/vllm-openai:stepfun37",
        "note": "Dedicated Step-3.7 image — preferred over the nightly pip wheel until support lands in vllm:latest."
      }
    }
  },
  "features": {
    "tool_calling": {
      "description": "Step-3.5 tool call parser with automatic tool choice",
      "args": [
        "--tool-call-parser",
        "step3p5",
        "--enable-auto-tool-choice"
      ]
    },
    "reasoning": {
      "description": "Step-3.5 reasoning parser for chain-of-thought extraction",
      "args": [
        "--reasoning-parser",
        "step3p5"
      ]
    },
    "spec_decoding": {
      "description": "3-layer Multi-Token Prediction speculative decoding (MTP-3)",
      "args": [
        "--speculative-config",
        "{\"method\": \"mtp\", \"num_speculative_tokens\": 3}"
      ]
    }
  },
  "opt_in_features": [
    "spec_decoding"
  ],
  "variants": {
    "default": {
      "precision": "bf16",
      "vram_minimum_gb": 475,
      "description": "Full precision BF16 — recommended on 8xH200/B200 with TP8+EP"
    },
    "fp8": {
      "model_id": "stepfun-ai/Step-3.7-Flash-FP8",
      "precision": "fp8",
      "vram_minimum_gb": 238,
      "description": "Native FP8 checkpoint — runs on 8xH200/B200 with TP8+EP",
      "json": "/stepfun-ai/Step-3.7-Flash-FP8.json"
    },
    "nvfp4": {
      "model_id": "stepfun-ai/Step-3.7-Flash-NVFP4",
      "precision": "nvfp4",
      "vram_minimum_gb": 119,
      "tp": 4,
      "extra_args": [
        "--quantization",
        "modelopt",
        "--kv-cache-dtype",
        "fp8",
        "--gpu-memory-utilization",
        "0.9",
        "--async-scheduling"
      ],
      "description": "NVFP4 quantized — Blackwell only; TP4+EP with FP8 KV cache",
      "json": "/stepfun-ai/Step-3.7-Flash-NVFP4.json"
    }
  },
  "hardware_overrides": {},
  "guide": "## Overview\n\n[Step-3.7-Flash](https://huggingface.co/stepfun-ai/Step-3.7-Flash) is a 198B-parameter\nsparse Mixture-of-Experts vision-language model from\n[StepFun](https://www.stepfun.com/company), pairing a 196B language backbone with a\n1.8B perception encoder. It activates ~11B parameters per token and supports a\n256k context window with three selectable reasoning levels (low / medium / high).\n\nKey highlights:\n\n- **Multimodal Understanding**: Native vision encoder for single and multi-image inputs alongside text\n- **Hybrid Attention Architecture**: Interleaves Sliding Window Attention (512-token window) and Global Attention at a 3:1 ratio\n- **Sparse MoE**: 11B active parameters out of 198B total\n- **Multi-Layer MTP**: 3-way Multi-Token Prediction (MTP-3) for low-latency reasoning chains\n\nAvailable precisions:\n\n- [stepfun-ai/Step-3.7-Flash](https://huggingface.co/stepfun-ai/Step-3.7-Flash) (BF16)\n- [stepfun-ai/Step-3.7-Flash-FP8](https://huggingface.co/stepfun-ai/Step-3.7-Flash-FP8)\n- [stepfun-ai/Step-3.7-Flash-NVFP4](https://huggingface.co/stepfun-ai/Step-3.7-Flash-NVFP4) (Blackwell only)\n\n## Prerequisites\n\n- **vLLM version:** nightly (the model registry hasn't shipped in a stable release yet)\n- **Hardware:** 8xH200/B200 for BF16 and FP8; 4xB200 for NVFP4\n\n### Install vLLM (nightly)\n\n```bash\nuv venv\nsource .venv/bin/activate\nuv pip install -U vllm --pre \\\n    --extra-index-url https://wheels.vllm.ai/nightly\n```\n\nOr via Docker:\n\n```bash\ndocker pull vllm/vllm-openai:stepfun37\n```\n\n## Launching the Server\n\n### BF16\n\n```bash\nvllm serve stepfun-ai/Step-3.7-Flash \\\n    --served-model-name step3p7-flash \\\n    --tensor-parallel-size 8 \\\n    --enable-expert-parallel \\\n    --disable-cascade-attn \\\n    --reasoning-parser step3p5 \\\n    --tool-call-parser step3p5 \\\n    --enable-auto-tool-choice \\\n    --speculative-config '{\"method\": \"mtp\", \"num_speculative_tokens\": 3}' \\\n    --trust-remote-code\n```\n\n### FP8\n\n```bash\nvllm serve stepfun-ai/Step-3.7-Flash-FP8 \\\n    --served-model-name step3p7-flash \\\n    --tensor-parallel-size 8 \\\n    --enable-expert-parallel \\\n    --disable-cascade-attn \\\n    --reasoning-parser step3p5 \\\n    --tool-call-parser step3p5 \\\n    --enable-auto-tool-choice \\\n    --speculative-config '{\"method\": \"mtp\", \"num_speculative_tokens\": 3}' \\\n    --trust-remote-code\n```\n\n### NVFP4 (Blackwell only)\n\nRequires modelopt quantization and FP8 KV cache alignment.\n\n```bash\nvllm serve stepfun-ai/Step-3.7-Flash-NVFP4 \\\n    --served-model-name step3p7 \\\n    --tensor-parallel-size 4 \\\n    --gpu-memory-utilization 0.9 \\\n    --enable-expert-parallel \\\n    --quantization modelopt \\\n    --kv-cache-dtype fp8 \\\n    --reasoning-parser step3p5 \\\n    --tool-call-parser step3p5 \\\n    --enable-auto-tool-choice \\\n    --async-scheduling \\\n    --trust-remote-code\n```\n\n### DGX Station Single-GPU\n\nThe [DGX Station](https://www.nvidia.com/en-us/products/workstations/dgx-station/)\nships a single GB300 Grace-Blackwell Ultra Superchip with 252 GB of HBM3e, so the\nFP8 and NVFP4 checkpoints both fit entirely in VRAM on one GPU (BF16 at ~475 GB does\nnot). Use the dedicated `vllm/vllm-openai:stepfun37` image and serve on a single GPU:\n\n```bash\nvllm serve stepfun-ai/Step-3.7-Flash-FP8 \\\n    --served-model-name step3p7-flash \\\n    --tensor-parallel-size 1 \\\n    --gpu-memory-utilization 0.95 \\\n    --kv-cache-dtype fp8 \\\n    --reasoning-parser step3p5 \\\n    --tool-call-parser step3p5 \\\n    --enable-auto-tool-choice \\\n    --trust-remote-code\n```\n\nSwap in `stepfun-ai/Step-3.7-Flash-NVFP4` to run the NVFP4 checkpoint instead. Or\nlaunch the prebuilt container directly:\n\n```bash\ndocker run -d --name vllm-server \\\n    --gpus all --ipc host \\\n    --ulimit memlock=-1 --ulimit stack=67108864 \\\n    -p 8000:8000 \\\n    -e HF_TOKEN=\"$HF_TOKEN\" \\\n    -v \"$HOME/.cache/huggingface/hub:/root/.cache/huggingface/hub\" \\\n    vllm/vllm-openai:stepfun37 \\\n    stepfun-ai/Step-3.7-Flash-FP8 \\\n      --gpu-memory-utilization 0.95 \\\n      --trust-remote-code \\\n      --reasoning-parser step3p5 \\\n      --enable-auto-tool-choice \\\n      --tool-call-parser step3p5 \\\n      --kv-cache-dtype fp8\n```\n\nSee the [NVIDIA build DGX Station instructions](https://build.nvidia.com/station/vllm/instructions)\nfor the full container setup.\n\n## Benchmarking\n\n```bash\nvllm bench serve \\\n    --backend vllm \\\n    --model stepfun-ai/Step-3.7-Flash \\\n    --endpoint /v1/completions \\\n    --dataset-name random \\\n    --random-input 2048 \\\n    --random-output 1024 \\\n    --max-concurrency 10 \\\n    --num-prompt 100\n```\n\n## Troubleshooting\n\n- **MoE kernel tuning:** See [tune-moe-kernel](https://github.com/vllm-project/recipes/blob/main/Qwen/Qwen3-Next.md#tune-moe-kernel) to tune Triton kernels for your hardware.\n- **NVFP4 + TP > 4:** The author recommends TP4+EP for NVFP4. Higher TP isn't validated.\n- **Cascade attention:** Always pass `--disable-cascade-attn` — the hybrid SWA/GA schedule is not compatible with cascade attention in vLLM.\n\n## References\n\n- [Model card](https://huggingface.co/stepfun-ai/Step-3.7-Flash)\n- [FP8 checkpoint](https://huggingface.co/stepfun-ai/Step-3.7-Flash-FP8)\n- [NVFP4 checkpoint](https://huggingface.co/stepfun-ai/Step-3.7-Flash-NVFP4)\n- [StepFun](https://www.stepfun.com/company)\n",
  "source": {
    "source": "vllm",
    "repo": "vllm-project/vllm",
    "ref": "main",
    "sha": "564ac0de1b228568cafcbfbe8ccaccc82d6023eb",
    "path": "recipes/models/stepfun-ai/Step-3.7-Flash.yaml",
    "edit_url": "https://github.com/vllm-project/vllm/edit/main/recipes/models/stepfun-ai/Step-3.7-Flash.yaml",
    "overlays": []
  }
}