EfficientSAM3 v0.4.0 — Stage 3 Fine-Tuned Models & Full PCS Release

What's New

Apologies for the Delay — Fine-Tuned Models Are Here!

We know it's been a while, but good things take time! We're happy to share that the fine-tuned EfficientSAM3 models are finally ready.

This release brings lightweight image encoders — EV-M, RV-M, and TV-M — trained end-to-end on 5% SA1B data with SACap labels for full Promptable Concept Segmentation (PCS) capabilities.

Model	Vision Encoder	Text Encoder	Decoder	Other	Total Params	vs ImageSAM3
EV-M	EfficientViT-B1 (22.2M)	MobileCLIP-S0 (42.5M)	21.0M	3.5M	89.2M	90% smaller
RV-M	RepViT-M1.1 (25.6M)	MobileCLIP-S0 (42.5M)	21.0M	3.5M	92.7M	89% smaller
TV-M	TinyViT-11M (28.3M)	MobileCLIP-S0 (42.5M)	21.0M	3.5M	95.3M	89% smaller

Note: ImageSAM3 comparison: Vision (463M) + Text (354M) + Decoder (30.3M) + Other (14.2M) = 861.5M

Grab the checkpoints on HuggingFace.

We've also given the README a clean-up — easier to navigate, clearer model descriptions, and better organized for contributors.

Three-Stage Progressive Distillation

EfficientSAM3 delivers a complete distillation pipeline:

• Stage 1: Compact encoder distillation on SA-1B (image) + Recap-DataComp-1B (text)
• Stage 3: End-to-end fine-tuning on SAM3 data for full PCS quality

EfficientSAM3 v0.3.0 — EfficientSAM3.1 & SAM3.1-LiteText Image Model Release

EfficientSAM3 — v3.0 Models & Hugging Face Integration (2026-04-13)

We’re excited to announce the release of the EfficientSAM3.1 and SAM3.1-LiteText models, alongside official Hugging Face integration and advanced data engine support!

What’s new

• New Models Released: EfficientSAM3.1 and SAM3.1-LiteText image models are now available on the stage1_sam3.1 branch.
• Hugging Face Integration: SAM3-LiteText has been officially merged into HuggingFace Transformers.
• Data Engine Support: Stage 3 data engine support is now live on the data_engine branch.
  

🏆 Community Heroes

A massive thank you to the community members who helped make this release possible:

• @NielsRogge, @yonigozlan, and the Hugging Face team: For integrating SAM3-LiteText into the Hugging Face API.
• @colinlin1982: For finding redundant original SAM3 weights in the efficient models and adding a trim script to reduce the model size. (Note: The new EfficientSAM3.1 models do not need to be trimmed).
• @clcl777: For adding multi-device support.

EfficientSAM3 v0.2.0 — SAM3-LiteText

EfficientSAM3 — SAM3-LiteText (2026-02-18)

We’re releasing SAM3-LiteText, a highly efficient text encoder variant for SAM3.

• What’s new

  • SAM3-LiteText released! Reduces text encoder parameters by up to 88% with similar performance to the original text encoder.
  • Paper Released: arXiv:2602.12173 (PDF)
  • Code: Available in the sam3_litetext branch.

• Get the weights

  • Download from Hugging Face: Simon7108528/EfficientSAM3/tree/main/sam3_litetext

• Quick start

  • Checkout the branch: git checkout sam3_litetext
  • Install: pip install -e ".[stage1]"
  • Run inference with the new lite text encoders.

EfficientSAM3 v0.1.1 — Stage 1+ Fine-Tuned (ft) Weights

EfficientSAM3 — Stage 1+ Fine-Tuned (ft) Weights (2026-01-11)

We’re releasing/refreshing Stage 1 geometry-prompt fine-tuned (ft) checkpoints for EfficientSAM3.

• What’s new

  • Image encoder ft weights (trained on 1% SA-1B with geometry-prompt fine-tuning)
  • Text encoder ft weights (fine-tuned on SA-Co Gold + Silver text annotations)
  • Updated download links in the Model Zoo (Google Drive + Hugging Face)

• Get the weights

  • See the Model Zoo in the repo README.md and the project page: https://simonzeng7108.github.io/efficientsam3/

• Quick start

  • Install: pip install -e ".[stage1]"
  • Inference examples are in the repo under sam3/efficientsam3_examples/