gaea-plugin-mage-vl

Image understanding for Gaea: Microsoft's Mage-VL 4B (microsoft/Mage-VL, Apache-2.0) on mage-vl-swift, serving MES image.describe. Add it by its repo URL:

gaea plugins add https://github.com/xocialize-code/gaea-plugin-mage-vl

What it does

• One model, mage-vl, in one lane: int8 (mlx-community/Mage-VL-8bit, about 5.4 GB resident), as pixel-agency-image-engine v0.9.1 serves it. Its greedy output is byte-identical to bf16 and to the PyTorch reference.
• One offer, image.describe (Image profile). It takes image and the text input question, and answers the value text. With no question, it asks "Describe this image in one or two sentences.", as v0.9.1 does.
• Any size in. Mage-VL resizes the picture to its own pixel cap, so the offer's limit (100 MP) is only a sanity bound.

In Gaea, the image normalizer asks for a caption, alt text and tags as JSON (in the caller's language), and parses the reply into the profile's description.

Build and check locally

scripts/build-plugin.sh

scripts/check-plugin.sh

check-plugin.sh describes a fixture photo over GPP with the image normalizer's question, with every .build folder hidden. A control then shows that the bare worker, outside the bundle, can't find MLX's metallib.

swift build -c release --product mage-vl-inproc && scripts/golden.sh <out-dir>

golden.sh describes three photos (two fixtures and a DIV2K validation photo) with the default question and the normalizer's, over GPP and in-process (mage-vl-inproc: MLXServeEngine, the same registration and request), and compares the texts. scripts/release.sh builds, checks, notarizes with the fleet's mlx-notary profile and publishes the GitHub release; see the demo plugin for signing and CI secrets.

Licence

The plugin is Apache-2.0 (Gaea G-21). Mage-VL's weights and mage-vl-swift's port are Apache-2.0. The fixture photos come from gaea-plugin-forge.