FocsleFo’c’sle
Fo’c’sle

SAM-3-Distilled

28%
by Fo’c’sle

Distilled SAM-3 for VLA stacks that need on-device segmentation prompts. 4× lighter than the reference.

SegmentationFocsle-ResearchINT8FP16samdistilledvla-companion
84K downloads 2.4K deploymentsUpdated Apr 19, 2028
Headline:14.2ms · NVIDIA Jetson Thor · MIXED
OverviewBenchmarks5Sim ResultsDeploy5FilesDiscussion23

Deploy SAM-3-Distilled

Pick a chip family. We hand you the artifacts (HEF, TRT engine, Core ML, ONNX) plus a one-click endpoint deploy. For private endpoints, on-prem deploy, or air-gapped distribution, see Enterprise.

NNVIDIA Jetson AGX Orin
# Build a TensorRT engine
$ focsle pull focsle/sam-3-distilled --target jetson-agx-orin
$ focsle build trt --plan sam-3-distilled.plan \
    --precision fp16 \
    --workspace 4G

# Run with TensorRT
import focsle.runtime as fr
m = fr.load("sam-3-distilled.plan", target="trt")
out = m.run(frame)

One-click endpoint

Spins up a managed endpoint in the closest region. Pro and above.

Or deploy yourself