AbstractFramework/fibo-8bit

FIBO 8-bit for MLX-Gen

This repository contains MLX-Gen saved weights for briaai/FIBO. It is designed for local Apple Silicon inference with mlx-gen.

This is an MLX-Gen/mflux saved-weight folder, not a Diffusers or Transformers from_pretrained() checkpoint.

License and access

This checkpoint is a quantized derivative of the gated briaai/FIBO source model and follows the source model's bria-fibo non-commercial license terms. Access to this derivative should require users to accept the upstream FIBO terms before downloading the files.

Capability

This package is for base FIBO text-to-image generation only.

FIBO Edit is a separate upstream model family and is not included in this package. Current MLX-Gen public capability discovery intentionally does not expose FIBO Edit through unified mlxgen generate until source-model and prepared-package edit validation pass.

Quantization

This is a mixed q8/BF16 FIBO package:

• q8 for quantizable FIBO transformer and text-encoder linears.
• BF16 for the VAE.
• BF16 for FIBO conditioning, timestep, caption-projection, normalization, and output-projection paths that are precision-sensitive in MLX-Gen.

Validation

Measured on an Apple M5 Max with 128 GB unified memory using mlx-gen 0.18.11.

Prompt file: validation/fibo_seed103_prompt.json

Settings: width=512, height=512, steps=8, guidance=5, seed=103.

Package	Disk	Time	Max RSS	MAE vs source
Source briaai/FIBO	23.8 GiB cache	15.13 s	24.13 GB	baseline
This q8/BF16 package	14.5 GiB	16.45 s	15.89 GB	8.81
q4/BF16 package	10.2 GiB	15.24 s	11.39 GB	17.56

Detailed validation notes: validation/fibo_base_q8_q4_seed103_metrics.md

Usage

python -m pip install -U mlx-gen

mlxgen download --model AbstractFramework/fibo-8bit

mlxgen generate \
  --model AbstractFramework/fibo-8bit \
  --prompt-file validation/fibo_seed103_prompt.json \
  --width 512 \
  --height 512 \
  --steps 8 \
  --guidance 5 \
  --seed 103 \
  --output fibo-q8.png

Attribution

FIBO is published by Bria.ai. MLX-Gen is based on mflux by Filip Strand and the original mflux contributors.

Quantized and contributed by @lpalbou.