gaea-plugin-moss-image

OpenMOSS's MOSS-Image (a 5B single-stream DiT on a Qwen3.5-9B text encoder and the FLUX.2 small decoder, Apache-2.0) as a Gaea research plugin. It is built on GaeaPluginKit over moss-image-swift 0.3.0, runs in its own process, and speaks the Gaea Plugin Protocol v1. Add it to Gaea by its repo URL:

gaea plugins add https://github.com/xocialize-code/gaea-plugin-moss-image

What it offers

Model	MES operation	Quality tiers	Notes
moss-image-sft ("MOSS-Image")	image.generate	standard, best	Long English and Chinese text (line breaks in the prompt are kept), and JSON layout prompts that place text and objects by bounding box (summary.caption + items, yxyx in pixels of the requested canvas).
moss-image-base ("MOSS-Image Base")	image.generate	best	The pretrained checkpoint SFT is tuned from: the same prompts and layout JSON, less curated composition. For LoRA work and comparison.

Every offer is named-only: it runs only when a request names the model. See Why named-only.

• Sampling. 50 Euler steps with CFG 3 (two forwards a step): about 2 minutes per 1024² image on an M5 Max. Negative prompts are honoured.
• Sizes. Sides in multiples of 16, each at least 256 px, at most 1456² in area (the package's MossImageEnvelope, where its footprints were measured). There is no side cap of its own. A larger area is scaled down keeping the aspect, then floored to multiples of 16, and reported size_snapped; a shorter side under 256 px is refused. The default is 1024². aspect with megapixels works too.
• count: 1–8 images, each with seed + i. A request without a seed gets one, reported in effective.
• Options, as MES extensions: x-moss.steps (1–100, default 50) and x-moss.guidance (1–20, default 3).
• Not offered: editing (MOSS-Image is text-to-image only), and LoRA adapters.

Lanes and memory

Each model has two lanes, each a DiT tier paired with a text-encoder tier, as mlx-image-runtime's catalog pairs them on a Mac (ModelCatalog.mossQuant and mossTextEncoderQuant, AB-D-0152). The figures are the package's declared footprints, as dist/manifest.json records them:

Lane	DiT	Encoder	Resident	Peak while running	Gaea schedules
bf16 (default)	bf16	bf16 (16.8 GB)	10.5 GB	16.7 GB	27.2 GB
int8	int8	int8 (9.4 GB)	5.8 GB	10.4 GB	16.2 GB

The encoder loads for each request and is released before the denoise loop, so the peak is the encode, at every size up to 1456² (moss-image-swift 0.3.0 decodes the VAE in row bands). In-app measurements: bf16/bf16 26.5 GB, int8/int8 15.7 GB at 1024² and at the cap. int8 quantizes the bf16 weights on load; its text error is within noise of bf16 (an OCR gate over 20 English and Chinese long-text prompts, AB-R-0572). The package declares no int4 tier: it fails the quality gate.

bf16 is the default. When it doesn't fit, Gaea loads the int8 lane and warns (lane_fallback, G-23); gaea models lane moss-image-sft int8 makes it the default. A request naming a lane (x-gaea.lane) never falls back.

Weights

Gaea sends its models root in configure (MLXEngine's store by default, gaea models root to change it). The weights come from the mlx-community mirrors, pinned by commit and byte-identical to OpenMOSS's upstream revisions:

Repo	Commit	What it gives
mlx-community/MOSS-Image-SFT-bf16	ea698e99312015505884559ebad94b4a5a0d0753	SFT's transformer/, and the shared text_encoder/, tokenizer/ and vae/
mlx-community/MOSS-Image-Base-bf16	56e5c1f98640356efce285957df9a883a53064fa	Base's transformer/

Both models read the Qwen3.5-9B encoder, tokenizer and VAE from the SFT mirror, so a store holding both keeps the 16.8 GB encoder once, and Base needs the SFT mirror's shared files too. The worker downloads a missing snapshot on prepare or materialize.

Build and check locally

scripts/build-plugin.sh

scripts/check-plugin.sh

check-plugin.sh generates one 256² image in 8 steps with MOSS-Image int8 (short on purpose) against the weights in MLXEngine's store (CHECK_MODELS_ROOT moves it), with every .build folder hidden so MLX can find its metallib only inside the bundle. A control then shows that the bare worker, outside the bundle, can't. scripts/release.sh builds, checks, notarizes with the fleet's mlx-notary profile, and publishes the GitHub release. See the demo plugin for signing, notarization and CI secrets.

moss-inproc (swift run -c release moss-inproc; its header lists the options) is not part of the plugin: it runs the same lane in-process through MLXServeEngine and builds the same request as the worker (the library's snapped and generateRequest), the reference that Gaea's runs are compared against byte for byte.

Why named-only

This is a research plugin (Gaea AB-T-0374), one of a family of opt-in plugins for models the fleet ports but doesn't default to. The catalog serves MOSS-Image as prod and preferred for nothing: at about 2 minutes per 1024² image it is for its text rendering and layout prompts, asked for by name. Gaea makes an offer named-only by itself only when the weights licence isn't permissive, so this plugin sets "named_only": true on every offer, and the normalizer never picks a MOSS model by default.

Licence

The plugin is Apache-2.0 (Gaea G-21) and ships no weights. MOSS-Image's weights (the DiT, the Qwen3.5-9B encoder and the FLUX.2 small decoder) are Apache-2.0 (OpenMOSS); shipping was approved on the tech report's data account (AB-D-0150). moss-image-swift is MIT.