# Microsoft Open-Sources 4B-Parameter Multimodal Model Family Mage

- Category: Flash (快讯)
- Published: 2026-08-19T11:58:00.000Z
- Source date: 2026-08-19
- HTML: https://www.88z.ai/en/news/flash-20260819-b46161ace8ab1c44
- Markdown: https://www.88z.ai/en/news/flash-20260819-b46161ace8ab1c44.md
- Other language: https://www.88z.ai/news/flash-20260819-b46161ace8ab1c44.md

Microsoft has released the open-source multimodal model family Mage, covering core capabilities in visual understanding and image generation. Mage-VL significantly reduces visual tokens via Codec-Native encoding, boosting video inference speed and enabling real-time streaming processing. Mage-Flow focuses on high-resolution image generation and editing, achieving rapid output on a single A100 GPU through efficient sampling in its Turbo version, while supporting diverse resolution requirements.

## At a glance

- Mage-VL uses Codec-Native encoding to cut visual tokens by 75% and boost video inference speed by 3.5x
- Mage-Flow Turbo requires only 4 sampling steps, generating images in 0.6 seconds on a single A100 GPU

## The update

Microsoft has officially open-sourced Mage, a multimodal model family with 4 billion parameters. The series primarily includes Mage-VL for streaming video and image understanding, and Mage-Flow for image generation and editing. Mage-VL introduces Codec-Native encoding technology, reducing visual tokens by 75 percent, increasing video inference speed by 3.5 times, and enabling real-time streaming understanding. Mage-Flow supports text-to-image generation and instruction editing at any native resolution between 512 and 2048. Its Turbo version requires only four sampling steps and can generate an image in 0.6 seconds on a single A100 GPU, demonstrating efficient generation performance.

## Sources

- [Mage](https://www.88z.ai/flash/cite/flash-20260819-b46161ace8ab1c44) (2026-08-19)

This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.