VisualInference/MOSS-Image-Base-MLX-8bit

MOSS-Image-Base for Apple Silicon (MLX, 8bit)

An MLX build of OpenMOSS-Team/MOSS-Image-Base, the pre-trained (not instruction-tuned) MOSS-Image checkpoint, for Macs with Apple Silicon. It uses the same quantization recipe as the verified SFT 8bit tier and runs with moss-image-mlx.

Base is the checkpoint OpenMOSS rendered their official sample gallery with, and the one to use for LoRA work and for reproducing those samples. For general use, the SFT tiers give more organised compositions.

• Peak memory: ~9.3–10.2 GB for 1–1.5 MP images.
• Download: 14.1 GB.
• Official seeds reproduce official samples, below.

Official sample vs this tier, from the official seed and settings

Official (OpenMOSS, CUDA, bf16) · This tier (MLX 8-bit, M4 Pro)

Sample	PSNR vs official	PSNR vs the port's own bf16	Time	Peak
#013, 768×1360, 650-token prompt	26.0 dB	27.7 dB	13.5 min	10.2 GB
#009, 1024×1536, CFG 5	19.6 dB	19.9 dB	20.1 min	9.3 GB

With this model, bf16 rounding alone moves an image about 23 dB, so these are differences in fine detail, not composition. See verification.

Quickstart

pip install git+https://github.com/felix-ab/moss-image-mlx
moss-mlx --model VisualInference/MOSS-Image-Base-MLX-8bit --parameters parameters.json -o sample.png     # an official example, exactly
moss-mlx "a misty harbor at dawn, tag style: best quality, highres" --model VisualInference/MOSS-Image-Base-MLX-8bit --guidance 4 -o harbor.png

Prompting: OpenMOSS recommends at least 50 steps (the default) and CFG 3–6 for Base, raising CFG for short prompts.

License and attribution

Apache-2.0, the same as the original weights. See NOTICE.

• MOSS-Image: © OpenMOSS.
• Text encoder: Qwen3.5-9B (Apache-2.0).
• VAE: FLUX.2 with the FLUX.2 small decoder by Black Forest Labs (Apache-2.0).

This is an independent conversion, not affiliated with or endorsed by OpenMOSS.