BAGEL¶
| Model | Source | Revision | Weight license | Parity (CPU, fp32) |
|---|---|---|---|---|
bagel-7b-mot |
ByteDance-Seed/BAGEL-7B-MoT | 5019f57 |
Apache-2.0 | Verified |
Catalog: bagel ·
Models: bagel-7b-mot · Task: t2i
Architecture. BAGEL is a unified 14B-parameter Mixture-of-Transformers with about 7B active parameters per token. Text and image-boundary tokens use the Qwen2 understanding expert while latent patches use a separate generation expert; both routes share one attention sequence. A rectified-flow head predicts 2×2 patches of 16-channel FLUX VAE latents. The published SigLIP vision branch is loaded strictly but is not executed for text-only generation.
Usage.
from imagehub import pipeline
with pipeline(
"t2i", backend="bagel-7b-mot", model="checkpoints/bagel-7b-mot"
) as pipe:
result = pipe("a cat, oil painting", seed=42)
result.images[0].save("cat.png")
Settings. The supported options are width and height (1024, multiples of 16 up to 1024),
steps (50), guidance (4), timestep_shift (3), cfg_interval ((0.4, 1.0)),
cfg_renorm_min (0) and seed. The release is bfloat16 and needs roughly 30 GB just for model
weights. Unrecognized settings raise an error before loading.
Validation. The reference suite uses the official InterleaveInferencer and
Bagel.generate_image. Its end-to-end checks use two flow timesteps at 64×64 for seeds 0 and 42;
these deliberately shortened runs validate ordering and determinism without attempting the
release's expensive 50-step 1024×1024 default. Real-checkpoint results are pending the dedicated
validation run.
Out of scope. Image understanding, image editing, interleaved multi-image generation, reasoning-token generation, quantized loading and TaylorSeer acceleration are not supported.
License. The code and weights are Apache-2.0. imagehub does not distribute weights.