README.md

---
library_name: mlx
license: apache-2.0
license_link: https://huggingface.co/OpenMOSS-Team/MOSS-Image-Base/blob/main/LICENSE
pipeline_tag: text-to-image
base_model: OpenMOSS-Team/MOSS-Image-Base
language: [en, zh]
tags: [mlx, safetensors, apple-silicon, text-to-image, diffusion, moss-image]
---

# mlx-community/MOSS-Image-Base-bf16

MLX (bf16) conversion of **[OpenMOSS-Team/MOSS-Image-Base](https://huggingface.co/OpenMOSS-Team/MOSS-Image-Base)** (Apache-2.0) for Apple Silicon. **MOSS-Image-Base** is the pretrained checkpoint that MOSS-Image-SFT is fine-tuned from: same architecture and prompts, with less curated composition. It is the starting point for LoRA training. For generation, prefer [MOSS-Image-SFT-bf16](https://huggingface.co/mlx-community/MOSS-Image-SFT-bf16).

## Model

MOSS-Image pairs a 5B single-stream DiT with the **Qwen3.5-9B** text encoder and the **FLUX.2 VAE** with BFL's small decoder. Its DiT design:

- 42 blocks with one shared adaLN modulation
- SiTU-GLU FFN
- 4-axis RoPE

## What it does well

- Long text, in English and Chinese. Line breaks in the prompt are kept.
- JSON layout prompts that place text and objects by bounding box: `summary.caption` plus `items`, with `bbox_format: "yxyx"` and `bbox_coord_space: "abs_centralized"` (pixels, centred on the canvas, same size as the request).

The snapshot is a standard diffusers tree (`transformer/ text_encoder/ tokenizer/ vae/ scheduler/`):

- The transformer and text encoder are bf16, the VAE fp32. The files are byte-identical to upstream revision `1ceede80`.
- The PyTorch custom pipeline (`pipeline_moss_image.py`) stays upstream.

## Sampling

| Setting | Value |
|---|---|
| Sampler | Euler flow matching |
| Steps | 50 |
| Shift | 5 (snapped to the training grid) |
| CFG | 3, with block 22 skipped in the negative branch |
| Negative prompt | empty |

## Parity (Swift/MLX port vs the PyTorch reference, fp32, TF32 off)

| Component | Result |
|---|---|
| Qwen3.5-9B encoder, final hidden state | relL2 4.8e-6 |
| DiT, one forward (12 timestep / prompt / skip-layer cases) | relL2 1.1e-6 – 5.3e-6 |
| Scheduler sigmas (4 – 50 steps) | bit-exact |
| FLUX.2 small decoder | relL2 1.2e-6 |
| bf16 end-to-end (8 steps) vs fp32 | inside PyTorch's own bf16 drift |

## Memory and speed (M5 Max, 1024², 50 steps)

| DiT tier | Resident | Peak | Denoise |
|---|---|---|---|
| bf16 | 10.3 GB | 26.2 GB | 113 s (2.27 s/step) |
| int8 (quantized at load) | 5.6 GB | 21.5 GB | about the same |

- Every step runs two transformer forwards (CFG).
- The peak is set by the bf16 text encoder (16.8 GB). The port loads it per request and releases it before denoising.
- Sizes from 256² to 1456² by area work; width and height must be multiples of 16.

## Use (Swift / MLXEngine)

```swift
import MLXMossImage
import MLXToolKit

let package = MossImageBasePackage(configuration: MossImageConfiguration(quant: .bf16))  // weights materialize on load
try await package.load()
let r = try await package.run(T2IRequest(prompt: "A hand-painted sign that reads OPEN LATE", width: 1024, height: 1024, seed: 42)) as! T2IResponse
```

Code: https://github.com/xocialize/moss-image-swift (Swift/MLX port, MIT).

## Training data (from the upstream technical report)

- **Sources:** about 3.2 billion image-text records from 16 public datasets. About 98% come from ReLAION, OmniCorpus and DataComp.
- **Filtering:** AI-generated images were filtered out.
- **Captions:** written by Gemma 4 and a Qwen model.
- **Rights:** the report makes no rights statement for the data.

Review the [technical report](https://github.com/OpenMOSS/MOSS-Image/blob/main/paper/MOSS_Image_Technical_Report.pdf) before you use the model.

## License

The weights are Apache-2.0, as released by OpenMOSS. Components:

- Text encoder: Qwen3.5-9B, Apache-2.0.
- VAE: FLUX.2 small decoder (Black Forest Labs), Apache-2.0.