Mage-VL 跑在 Mac 上:微软视觉大模型的本地化部署,图片视频实时流式注释

Mage-VL on a Mac: Running Microsoft's Vision Model Locally with Real-Time Image and Video Annotation

Tech-News #Mage-VL#微软#Apple Silicon#本地推理#视频理解#多模态#MPS#Mycelium
更新于
🇨🇳 中文

by Mycelium Protocol


微软的 Mage 研究团队做了一个视觉语言模型 Mage-VL,能看懂图片和视频,还能对视频内容做实时的分段流式注释——“边看边说”,而不是等视频处理完再输出。

问题是,官方仓库(microsoft/Mage)面向的是研究环境:CUDA 优先,命令行推理,需要自己组装完整的研究工具链。

mage-vl-local-mac(karlazx)解决的就是这个问题:让 Apple Silicon 用户双击两次就能跑起来,在浏览器里用。


它是什么

mage-vl-local-mac 是 Mage-VL 的 macOS 部署和体验层,不是微软官方项目。社区项目,MIT 授权,2026 年 8 月 1 日发布。

架构:

React UI(localhost:3000)
    ↓
FastAPI(localhost:8000)→ 单用户推理队列 → 视觉预处理
                                              ├── 图片缩放(quick/balanced/原始分辨率)
                                              ├── 视频均匀采帧
                                              └── DCVC-RT 神经编解码器(MPS → CPU fallback)
                                                          ↓
                                               Mage-VL(PyTorch + Apple MPS)
                                                          ↓
                                               SSE 流式输出 → 浏览器

三个核心能力

图片问答

上传图片,提问,Mage-VL 回答。支持三种视觉预算模式:

  • quick:快速,低分辨率,省内存
  • balanced:平衡,适合大多数场景
  • original:原始分辨率,最精确,内存消耗最高

视频问答

上传视频(支持到 150 秒),两种处理路径:

  • 均匀采帧:传统方式,从视频均匀抽帧作为输入
  • DCVC-RT 神经编解码器:把视频解码为神经”画布”(canvas),是 Mage 研究的核心路径,比传统 H.264/HEVC 帧更适合模型理解

DCVC-RT 优先在 MPS 上运行,不支持时自动退回 CPU。

实时分段流式注释(Proactive Streaming)

这是 Mage-VL 的特色能力。对于长视频,模型不会等全部处理完再输出,而是:

  1. 处理第一段视频 → 立刻输出这段的注释
  2. 处理第二段视频,携带第一段的注释作为上下文 → 输出第二段注释
  3. ……以此类推

类似直播主播在”边看边讲”。浏览器通过 SSE(Server-Sent Events)实时接收每个文字 token。

你还可以在运行中编辑”实时注释指令”,比如让模型专注于某类内容,也有重复文字抑制机制,防止连续段落输出相同的句子。


一键安装

# 方法 1:下载 Release ZIP,解压后双击
# → setup.command → start.command → 打开 localhost:3000

# 方法 2:克隆源码
git clone https://github.com/karlazx/mage-vl-local-mac.git
cd mage-vl-local-mac
./setup.command   # 安装 Python 3.12、Node.js 22、FFmpeg、模型权重
./start.command   # 启动服务

setup.command 会自动通过 Homebrew 安装依赖,创建隔离的 Python 环境,构建前端,然后从 Hugging Face 下载 microsoft/Mage-VL 权重(约 25 GB)。模型权重不包含在仓库里,首次运行自动下载,支持断点续传。


硬件要求

  • Apple Silicon Mac(M1/M2/M3/M4 均可)
  • 推荐 32 GB 统一内存(完整模型权重 + 视频处理需要)
  • 约 25 GB 可用磁盘
  • 低内存机器在 quick/balanced 图片模式下可能可用,视频任务可能超出内存限制

MacBook Air 无风扇,长视频推理可能触发热降频。


隐私

所有推理本地进行。用户上传的媒体文件存在本地 runtime/ 目录,不发送到任何云端 API 或这个项目的服务器。临时文件在过期后自动清理。


与官方 microsoft/Mage 的关系

方面microsoft/Mage(官方)mage-vl-local-mac(本项目)
目标研究代码,CUDA 环境macOS 一键部署,浏览器 UI
硬件路径CUDA 优先Apple MPS,FP16/BF16
视频编解码传统 H.264/HEVC + DCVC-RTDCVC-RT + 均匀采帧
训练 / Mage-Flow包含不包含

如果你做研究、需要训练、或者遇到模型问题,应该去官方仓库报 issue。这个项目专注的是”Mac 上好用”这件事。


本地 API

服务启动后暴露一个 REST + SSE API,绑定在 127.0.0.1:

Endpoint用途
GET /api/status模型、MPS、DCVC、队列状态
POST /api/jobs创建图片/视频/流式注释任务
GET /api/jobs/{id}/eventsSSE 流,获取 token/时间线/最终结果
GET /api/artifacts/{id}/{file}读取生成的帧和画布

单用户推理队列,同时只处理一个任务,适合统一内存的单机场景。


为什么现在值得关注

Mage-VL 的视频理解能力,特别是 DCVC-RT 神经编解码器路径和实时分段流式注释,在多模态模型里不多见——大多数本地视频理解工具还在用传统采帧的方式。

karlazx 把这个研究能力包成了一个对普通 Mac 用户可以双击运行的工具,这个”最后一公里”的工作本身就有价值。

仓库:github.com/karlazx/mage-vl-local-mac


Mycelium Protocol — 追踪 AI 系统的底层演化


关于 Mycelium

菌丝协议。持续追踪 AI 工具、系统和实验的内容节点。


🇬🇧 English

Mage-VL on Your Mac: Microsoft’s Vision Model, One-Click Local Deployment

by Mycelium Protocol

Microsoft’s Mage research team built Mage-VL, a vision-language model that understands images and videos and can produce real-time segment-by-segment streaming commentary — “speaking while watching” rather than waiting until the full video is processed.

The problem: the official repository (microsoft/Mage) targets research environments — CUDA-first, command-line inference, requiring the full research toolchain to be assembled manually.

mage-vl-local-mac (karlazx) solves that: Apple Silicon users can double-click twice and use it in a browser. Community project, MIT license, released August 2026.

Architecture

FastAPI backend at localhost:8000, React frontend at localhost:3000. A single-user inference queue manages jobs; visual preprocessing dispatches to three paths: image resize presets, uniform frame sampling, or DCVC-RT neural codec canvases. All routes feed into Mage-VL via PyTorch + Apple MPS. Results stream to the browser over Server-Sent Events.

Three Core Capabilities

Image Q&A — Upload an image, ask a question, get an answer. Three visual budget modes: quick (low-res, low memory), balanced (general purpose), and original-resolution (maximum detail, highest memory).

Video Q&A — Upload a video (up to 150 seconds). Two processing paths:

  • Uniform frame sampling: traditional approach, uniform frame extraction
  • DCVC-RT neural codec: decodes video into neural “canvases” — the core research path in Mage, more semantically appropriate for the model than raw H.264/HEVC frames. Runs on MPS; falls back to CPU if unsupported

Proactive Streaming Commentary — The distinctive Mage-VL capability. For long videos, the model doesn’t wait until all processing is complete:

  1. Process segment 1 → immediately publish commentary for segment 1
  2. Process segment 2, using segment 1’s commentary as context → publish segment 2
  3. …continuing through the video

Think of a commentator narrating live footage. You can edit the live-commentary instruction during playback — redirect the model’s focus mid-video — and repeated-text suppression prevents duplicate sentences across consecutive segments.

One-Click Setup

git clone https://github.com/karlazx/mage-vl-local-mac.git
cd mage-vl-local-mac
./setup.command   # installs Python 3.12, Node.js 22, FFmpeg; downloads ~25 GB model weights
./start.command   # starts the service
# open http://localhost:3000

Or download the Release ZIP, extract, and double-click setup.command. The model weights (Microsoft/Mage-VL from Hugging Face) download automatically with resumable support; they’re not bundled in the repository.

Requirements: Apple Silicon, macOS with Homebrew, 32 GB unified memory recommended, ~25 GB free disk. Works on 8–16 GB machines for image quick/balanced modes; long video jobs may exceed available memory.

Privacy

All inference is local. Media uploaded through the browser stays under the local runtime/ directory and is never sent to any cloud API or this project’s servers. Temporary artifacts are cleaned up when they go stale.

vs. Official microsoft/Mage

The official repository is the canonical source for Mage research, training, Mage-Flow, and the traditional codec pipeline. Use the official repo for upstream issues, model research, and training. Use this project when the priority is a convenient native Mac deployment and a usable browser-based product surface.

Why It Matters Now

Most local video understanding tools still rely on traditional frame sampling. Mage-VL’s DCVC-RT neural codec path and real-time proactive streaming are less common in local deployments — they come from the research side where inference environments typically require CUDA. karlazx’s “last-mile” work — packaging this into a double-click Mac app — makes that research capability accessible to a much wider set of users.

Repository: github.com/karlazx/mage-vl-local-mac


Mycelium Protocol — tracking the deep evolution of AI systems

© 2026 Mycelium Protocol. All rights reserved.

💬 评论与讨论

使用 GitHub 账号登录后发表评论

关于本站 · 免责声明

🍄 Mushroom Research Blog 是非营利、免费公开的个人科技观察博客与公众号 XStack18,不接受商业合作、不代表任何企业或机构立场,也不谋求商业利益。我们以个人视角客观中立地记录和分析 AI、Web3 等领域的最新模型发布与技术动态——不止转述新闻标题或二手信息,而是给出有独立思考的深入分析,希望帮更多人获得有价值的一手科技认知。

⚠️ 文中介绍的开源代码与模型,仅供学习交流与技术借鉴。它们大多仍处于早期阶段,有待进一步研究和验证,请勿直接用于工作或生产环境;如需采用,请先自行充分测试,并核实其许可证与安全性。
Open-source code and models featured here are shared for learning and reference only. Most are early-stage and still need further study and verification — please don't use them directly in your work or in production. Test them thoroughly and check their licenses and security first.

  1. 本站文章均为作者基于公开信息的个人研究与观点整理,不代表文中提及的任何公司、产品、模型的官方立场,未与其构成商业关联或合作关系。
  2. 科技行业信息更新极快,我们尽力保证内容准确、及时,但不对完整性、实时性做绝对保证,具体请以相关企业/项目官方公告为准。
  3. 文中引用的第三方商标、产品名称、图片、数据等版权归原权利人所有,我们会尽量注明来源;如你认为存在版权疑问或侵权,请通过下方邮箱联系我们,收到通知后会尽快核实处理(更正、加注来源或删除)。
  4. 文章内容仅为技术科普与个人观点,不构成投资、法律或其他专业建议,据此进行任何决策的后果需自行判断和承担。

📮 侵权 / 勘误 / 合作咨询:hello@mushroom.cv