You need to agree to share your Hugging Face information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

bf16 copy of Meta's SAM 3, distributed under the SAM License (see LICENSE). By requesting access you agree to its terms.

Log in or Sign Up to review the conditions and access this model content.

SAM 3 bf16

facebook/sam3 cast from fp32 to bf16.

size
fp32 (original) 3.44 GB
bf16 1.72 GB

Floating-point tensors cast to bf16, everything else unchanged (configs, tokenizer, integer tensors). Conversion script: sam3-convert.py.

Mask IoU vs fp32: ≥ 0.999 on the tested images.

Examples

bf16 masks, text prompts car, fox, person.

import torch
from transformers import Sam3Model, Sam3Processor

model = Sam3Model.from_pretrained("cukrus/sam3-bf16", torch_dtype=torch.bfloat16).to("cuda")
processor = Sam3Processor.from_pretrained("cukrus/sam3-bf16")

Jetson Orin Nano 8GB: load 3 s, inference ~2 s, +0.2 s per extra text phrase.

License

SAM License, see LICENSE.

Downloads last month
-
Safetensors
Model size
0.9B params
Tensor type
BF16
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for cukrus/sam3-bf16

Base model

facebook/sam3
Quantized
(25)
this model