Models 2026-11: SAM 3

SAM 3 (facebook/sam3, revision 3c879f39826c281e95690f02c7821c4de09afae7) exported to ONNX for rascal-tools' sam3 segmenter (end to end: text prompt to instances and masks; three graphs, the CLIP tokenizer). Install with rascal tools get sam3. Licence: the SAM License (Meta), LICENSE.sam3.txt, which goes with every copy; this export is a derivative of SAM Materials and is distributed under it, and its use restrictions bind whoever uses it. Gates: graph parity (mask logits within 4.1e-3), instance parity, segment-01 corpus 20/20 at IoU >= 0.999987. See tools/python/MODEL.md.

Models 2026-11: HQ-SAM ViT-L

HQ-SAM ViT-L (lkeab/hq-sam sam_hq_vit_l.pth, revision 09b02a333b37772133eff3997bdba997867374b7, Apache-2.0) exported to ONNX for rascal-tools' hqsam segmenter (OWLv2 + HQ-SAM, one outline per box, hq_token_only). Install with rascal tools get hqsam. Licence: LICENSE.sam-hq-vit-l.txt (Apache-2.0). Gates: export parity plane IoU 1.0; instance parity; segment-01 corpus 20/20 at IoU >= 0.999996. See tools/python/MODEL.md.

Models 2026-11: EVF-SAM2

EVF-SAM2 (code hustvl/EVF-SAM at 9935dc52571678b806830821b3ddeacb30e7af07; weights YxZhang/evf-sam2-multitask, revision df3644009016a4f62d13c02f3f8fb6c9bbf94310, Apache-2.0) exported to ONNX for rascal-tools' evfsam2 segmenter (BEIT-3 Large + SAM 2.1 Large, text to mask, one mask per query). Two graphs: evf-sam2.image.onnx, and evf-sam2.prompted.onnx with its weights in two external-data parts (evf-sam2.prompted.0.data, evf-sam2.prompted.1.data); the XLM-R tokenizer as evf-sam2.tokenizer.json. Install with rascal tools get evfsam2 (3.6 GB). Licence: LICENSE.evf-sam2.txt (Apache-2.0). Gates: export parity plane IoU 1.0; instance parity; segment-01 corpus 20/20 at IoU >= 0.999989. See tools/python/MODEL.md.

models-2026-11-birefnet: BiRefNet

BiRefNet (ZhengPeng7/BiRefNet at e2bf8e4460fc8fa32bba5ea4d94b3233d367b0e4, MIT) exported to ONNX (dynamo exporter, opset 19 for DeformConv): birefnet.onnx, input_image [1,3,1024,1024] to output_image logits, with the maker's MIT licence as LICENSE.birefnet.txt. Installed with rascal-tools get birefnet beside OWLv2 (SPEC.md "Choosable segmenters", decision 15). Export gate passed 2026-11-09: instance parity against the maker's PyTorch model, and the segment-01 corpus at IoU 1.000000 on 20 of 20 images.

models-2026-11-tiny: Tiny's licence files

The licence texts that go with Tiny's ONNX exports (OWLv2 base patch 16 and SAM 2.1 Hiera-Tiny, both Apache-2.0), installed beside the graphs by rascal-tools get tiny (SPEC.md "Choosable segmenters", decision 15). The graphs themselves stay in models-2026-10 and models-2026-11.

Model files (2026-11)

The SAM 2.1 hiera-tiny mask decoder re-exported to ONNX (opset 18) so segment can pick an object's whole outline and retry with points (PLAN.md M13, issue #14): it takes the box as its two corner points plus up to eight positive points and returns three candidate outlines with the decoder's predicted IoU of each. Same SAM revision (de431c4), Apache-2.0; the encoder and OWLv2 are unchanged and stay in models-2026-10. Interface and parity (every outline IoU 1.000 against PyTorch, box alone and with 1 and 3 points) in tools/python/MODEL.md; provenance in export.json.

rascal-tools fetch downloads it and checks it against tools/python/src/rascal_tools/artifacts.toml; release bundles carry it under models/.

file	sha256
sam2.1-hiera-tiny.decoder3.onnx	5e69b63545f044fcc000f12c9cefc52801091d1b07d553ad69af8505cd88809e

v0.1.1

Download rascal-mask 0.1.1

File	Platform	Checksum
rascal-mask-aarch64-apple-darwin.tar.xz	Apple Silicon macOS	checksum
rascal-mask-aarch64-unknown-linux-gnu.tar.xz	ARM64 Linux	checksum
rascal-mask-x86_64-unknown-linux-gnu.tar.xz	x64 Linux	checksum

v0.1.0

Download rascal-mask 0.1.0

File	Platform	Checksum
rascal-mask-aarch64-apple-darwin.tar.xz	Apple Silicon macOS	checksum
rascal-mask-aarch64-unknown-linux-gnu.tar.xz	ARM64 Linux	checksum
rascal-mask-x86_64-unknown-linux-gnu.tar.xz	x64 Linux	checksum

Model files (2026-10)

The model files rascal's segmenter pack uses (segment, the remove_prompt recipe): OWLv2 base patch16 for text-prompted detection and SAM 2.1 hiera-tiny for the masks, exported to ONNX (opset 18). Both are Apache-2.0; provenance, revisions and the export's parity checks are in export.json and tools/python/MODEL.md.

rascal tools setup (and rascal-tools fetch) downloads these and checks each against the SHA-256 in tools/python/src/rascal_tools/artifacts.toml; the release bundles carry them under models/.

file	sha256
owlv2-base-patch16.onnx	11a8711e6d8d63f7bf8bfcc0e62d5d4e00fa8b6096dbec82ee2fbbff61389767
owlv2-base-patch16.tokenizer.json	fbf857dd4099afc6da394d17cba2f22c5d01eb6d87bb74967c90634f3bcf4e82
sam2.1-hiera-tiny.encoder.onnx	ed386472cf112f10464cdad2d2c786e171306cb31aeb2096c6a1da8789c35073
sam2.1-hiera-tiny.decoder.onnx	53c6457d30ffca641fab504bf647ccb7d158c60e5d6c3ca810cf735c50e8cb59