README.md

---
license: other
license_name: sapiens2-license
license_link: https://github.com/facebookresearch/sapiens2/blob/main/LICENSE.md
pipeline_tag: keypoint-detection
library_name: mlx
base_model: facebook/sapiens2-pose-0.4b
tags:
- sapiens
- sapiens2
- human-centric
- pose
- mlx
- bf16
---
# mlx-community/sapiens2-pose-0.4b-bf16

bf16 MLX conversion of [`facebook/sapiens2-pose-0.4b`](https://huggingface.co/facebook/sapiens2-pose-0.4b) (Meta's
Sapiens2, ICLR 2026): 308-keypoint top-down pose heatmaps. Converted with mlx-vlm
0.7.0; the original float32 checkpoint is 2x this size.

What is in `model.safetensors` (0.85 GB):

- every parameter in bfloat16 (the reference runs inference in bf16 mixed
  precision);
- the q/k/v projections merged into one `wqkv` tensor per block, as the
  mlx-vlm Sapiens2 model expects.

Refer to the [original model card](https://huggingface.co/facebook/sapiens2-pose-0.4b) for the
model description, intended use and license.

## Use with mlx-vlm

```bash
pip install -U mlx-vlm
```

```python
from mlx_vlm import load
from mlx_vlm.models.sapiens2.generate import Sapiens2Predictor, read_image

model, _ = load("mlx-community/sapiens2-pose-0.4b-bf16")
predictor = Sapiens2Predictor(model)
# boxes: (N, 4) xyxy person boxes from a detector; defaults to the full image.
output = predictor.infer(read_image("image.jpg"), boxes=boxes, flip_test=False)
keypoints, scores = output["keypoints"], output["scores"]  # (N, 308, 2), (N, 308)
```

Outputs are numpy arrays at the input resolution (dense tasks) or in
source-image pixel coordinates (pose). See the mlx-vlm Sapiens2 README for
preprocessing details and the per-task output keys.