sam3

SAM 3 instance segmentation from text or point prompts (segment_text, segment_point; inside pick and place)

• Kind: segmenter
• Protocol: http
• Default port: 8810
• VRAM: 7.2 GB
• Load time: about 60 s
• Skills that use a segmenter: segmentation, manipulation, eef, reflexes
• Configurations that name it: default, analytic-manipulation, no-handlers, no-reactive, no-verifier, r1, r2, r3, reflexes, with-task-description

Weights

harness fetch sam3 --deploy <machine>

Directory	Source	From
<weights>/sam3	modelscope_model	facebook/sam3

Notes

The official sam3 package (CaP-X fork) installed editable in the project environment; sam3.pt is loaded with weights_only=True.

Manifest

kind = "segmenter"
summary = "SAM 3 instance segmentation from text or point prompts (segment_text, segment_point; inside pick and place)"
protocol = "http"
command = ["{python}", "-u", "services/servers/serve_sam3.py", "--checkpoint", "{weights}/sam3/sam3.pt",
           "--port", "{port}"]
port = 8810
# Measured on a 3090: 7.2 GB.
vram_gb = 7.2
load_seconds = 60
client = { url = "{url}" }
needs_files = ["{weights}/sam3/sam3.pt"]
notes = "The official sam3 package (CaP-X fork) installed editable in the project environment; sam3.pt is loaded with weights_only=True."

[[weights]]
dir = "sam3"
modelscope_model = "facebook/sam3"