Mage-VL-Cowork — CMF

Microsoft Mage-VL-Cowork — a 4B vision-language model for agents that work a computer through the screen and the terminal (66.0 % on OSWorld-Verified): the OneVision vision encoder over screenshots at native resolution on a Qwen3-4B language model that reasons in <think> and calls computer_use / bash tools — packaged as single CMF files. They run with cortiq, a Rust inference engine with no Python and no ML framework: NVIDIA, AMD and Intel GPUs through Vulkan/DX12, Apple silicon through Metal, and the CPU.

English · Русский · 中文

Quick start

cargo install cortiq-cli
hf download infosave/Mage-VL-Cowork-cmf mage-vl-cowork-q8_2f.cmf --local-dir .
cortiq run mage-vl-cowork-q8_2f.cmf --image screenshot.png --prompt "Task: open the Downloads folder." --max-tokens 1024

Use cortiq 0.8.51 or later; prebuilt binaries for Linux, macOS and Windows are on the releases page. --image takes a path, a data URL or an HTTP(S) URL and may be repeated (the images come before the text). The model thinks before it acts; the thinking counts toward --max-tokens. For agent use, pass the release's system prompt and tool schema (MAGE_HYBRID_SYSTEM_PROMPT on the upstream card) with --system or through the server; coordinates are on a 0–999 grid.

Files

Perplexity over the first 2048 tokens of wikitext-2; the original in Transformers scores 121.8 (an agent model, not a language model of prose).

The vision encoder and the merger stay 16-bit in both files and run in float32: the preprocessing and the rows handed to the language model match Transformers float32 (cosine 1.000000 for every row on the CPU, above 0.999 with the GPU's matrix units). Greedy = original: 48 greedy tokens of an image description against Transformers 4.57.3 float32.

Performance

cortiq 0.8.51, NVIDIA A40 (48 GB, Vulkan), cortiq bench:

A 2206×1048 screenshot is 9108 patches and 2277 language tokens; the vision encoder takes 4.8 s for it on the A40 (GEMMs and attention on the GPU). A running cortiq serve answers a 512×512 image in 1.7 s to the first token.

Apple M4 (Metal): q8_2f decodes 17.4 tok/s and reads a 2k-token prompt at 147 tok/s (its greedy image description is word for word the Transformers one), q4tp 36 and 295 tok/s — on a Mac with 16–24 GB the 4-bit file is the practical choice. The vision encoder takes 0.69 s for 512×512 and 10.8 s for the 2206×1048 screenshot.

Usage

cortiq run mage-vl-cowork-q8_2f.cmf --system "$(cat mage_system_prompt.txt)" \
  --image screen.png --prompt "Task: merge cells A1:D1 in budget.xlsx and save it."
cortiq run mage-vl-cowork-q8_2f.cmf --image a.png --image b.png --prompt "What changed on the screen?"
cortiq run mage-vl-cowork-q8_2f.cmf --image screen.png --image-max-pixels 1000000 --prompt "Find the search box."

--image-max-pixels N caps the screenshot resolution before patching (fewer tokens, a faster encoder); the release's processor allows up to 4 M pixels.

OpenAI-compatible server

cortiq serve mage-vl-cowork-q8_2f.cmf --port 8080

curl http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "mage-vl", "max_tokens": 1024,
       "messages": [{"role": "system", "content": "<MAGE_HYBRID_SYSTEM_PROMPT>"},
                    {"role": "user", "content": [
         {"type": "image_url", "image_url": {"url": "data:image/png;base64,..."}},
         {"type": "text", "text": "Task: close the dialog."}]}]}'

Image parts take an HTTP(S) URL, a data: URL or a local path; a malformed image is an HTTP 400. Tool calls come back in the model's <tool_call> JSON.

Hardware

NVIDIA / AMD / Intel (Vulkan, DX12): q8_2f needs about 6 GB of GPU memory, q4tp about 4 GB; on headless Linux install libvulkan1 libglvnd0 libegl1 libgl1 libglx0 and set XDG_RUNTIME_DIR=/tmp. Apple silicon: Metal is the default. CPU only: CMF_GPU=0.

Verify the download

sha256sum -c mage-vl-cowork-q8_2f.cmf.sha256

ca71b28811318523ff0b9742c0c1f0be5590f7bcfc10815cb0054f92dd170b88  mage-vl-cowork-q8_2f.cmf
aedc066d74bdb0f570541b2c53a74452b5f99e3fc39b364b223a8f0b351ead1f  mage-vl-cowork-q4tp.cmf

Reproduce

cortiq convert --model microsoft/Mage-VL-Cowork --quant q8_2f \
  --output mage-vl-cowork-q8_2f.cmf --also-output q4tp=mage-vl-cowork-q4tp.cmf

Reference and parity scripts: scripts/four.

License

Apache-2.0, as the original release; the weights were converted to CMF and the language model quantized (q8_2f, q4tp) by cortiq.

---

Документация на русском

Mage-VL-Cowork от Microsoft — визуально-языковая модель на 4 млрд параметров для агентов, работающих с компьютером через экран и терминал (66,0 % на OSWorld-Verified): энкодер OneVision по скриншотам в исходном разрешении поверх языковой Qwen3-4B, которая рассуждает в <think> и вызывает инструменты computer_use / bash. Здесь — файлы CMF для движка cortiq на Rust без Python: Vulkan/DX12, Metal, CPU.

cargo install cortiq-cli
hf download infosave/Mage-VL-Cowork-cmf mage-vl-cowork-q8_2f.cmf --local-dir .
cortiq run mage-vl-cowork-q8_2f.cmf --image screenshot.png --prompt "Task: open the Downloads folder." --max-tokens 1024

Нужен cortiq 0.8.51 или новее. | файл | языковая модель | энкодер | размер | перплексия wikitext-2 | назначение | |---|---|---|---:|---:|---| | mage-vl-cowork-q8_2f.cmf | 8 бит | 16 бит | 5.10 ГБ | 125.6 | по умолчанию | | mage-vl-cowork-q4tp.cmf | 4 бита | 16 бит | 2.98 ГБ | 128.7 | минимум памяти |

Энкодер и мерджер хранятся в 16 битах и считаются в float32: строки, передаваемые языковой модели, совпадают с Transformers float32 (косинус 1.000000 на CPU), 48 жадных токенов описания изображения у q8_2f — тоже. На NVIDIA A40 (Vulkan) q8_2f генерирует 66.8 ток/с, q4tp — 90.2; скриншот 2206×1048 (2277 токенов) энкодер обрабатывает за 4.8 с. На Apple M4 (Metal) q8_2f генерирует 17.4 ток/с, q4tp — 36 (на Mac с 16–24 ГБ удобнее q4tp). Агенту нужен системный промпт и схема инструментов релиза (MAGE_HYBRID_SYSTEM_PROMPT в исходной карточке) через --system или сервер; координаты — в сетке 0–999. cortiq serve принимает изображения в формате OpenAI (image_url: URL, data: или локальный путь). Лицензия Apache-2.0, как у оригинала.

---

中文说明

Mage-VL-Cowork 是 Microsoft 的 40 亿参数 视觉语言模型，面向通过屏幕与终端操作电脑的智能体（OSWorld-Verified 66.0%）：OneVision 视觉编码器以原生分辨率处理截图，语言部分为 Qwen3-4B，在 <think> 中推理并调用 computer_use / bash 工具。本仓库提供由 Rust 推理引擎 cortiq（无 Python）加载的 CMF 文件，支持 Vulkan/DX12、Metal 与 CPU。

cargo install cortiq-cli
hf download infosave/Mage-VL-Cowork-cmf mage-vl-cowork-q8_2f.cmf --local-dir .
cortiq run mage-vl-cowork-q8_2f.cmf --image screenshot.png --prompt "Task: open the Downloads folder." --max-tokens 1024

需要 cortiq 0.8.51 或更高版本。| 文件 | 语言模型 | 视觉编码器 | 大小 | wikitext-2 困惑度 | 用途 | |---|---|---|---:|---:|---| | mage-vl-cowork-q8_2f.cmf | 8 位 | 16 位 | 5.10 GB | 125.6 | 默认 | | mage-vl-cowork-q4tp.cmf | 4 位 | 16 位 | 2.98 GB | 128.7 | 最小内存占用 |

视觉编码器与融合层以 16 位存储、按 float32 计算：交给语言模型的特征与 Transformers float32 一致（CPU 上余弦 1.000000），q8_2f 的 48 个贪心解码 token 也与原模型一致。 NVIDIA A40（Vulkan）上 q8_2f 生成 66.8 tok/s，q4tp 90.2 tok/s；2206×1048 截图 （2277 个 token）编码耗时 4.8 s。Apple M4（Metal）上 q8_2f 生成 17.4 tok/s，q4tp 36 tok/s（16–24 GB 内存的 Mac 建议用 q4tp）。智能体用法需 通过 --system 或服务端传入发布方的系统提示与工具定义（见 原始模型卡 中的 MAGE_HYBRID_SYSTEM_PROMPT），坐标为 0–999 网格。cortiq serve 接受 OpenAI 格式的 图像（image_url：URL、data: 或本地路径）。许可证与原模型相同，为 Apache-2.0。