SAM 3.1 is the current checkpoint in Meta's concept segmentation line, out on 27 March 2026. The weights are free on Hugging Face, behind a license you have to accept: the SAM License, with trade-control terms ruling out military and warfare use, nuclear applications, espionage, and weapons work. Self-hosting also means Python 3.12 or newer, PyTorch 2.7 or newer, a CUDA 12.6 GPU, and cloning Meta's repository, since there is no Transformers integration to install instead. In your browser it runs on CNAPS Studio, included in the basic plan.
The interface is the one you know if you have used SAM 3: type a short phrase and every instance that matches comes back masked. What is new in CNAPS Studio is a checkbox called Merge to binary mask, which flattens all the per-instance masks into one black-and-white mask for the whole concept. That switch is the difference between an annotation output and a usable one, because a blur-out wants one mask covering every face in the frame, not eleven separate ones to combine.
The cost is time. Expect roughly 10 to 30 seconds per concept, so a batch of images against several phrases is not interactive work. Five segmentation models sit in the CNAPS Studio catalog, and this is the newest of the three you can talk to.
What it actually is
The architecture is inherited: a vision encoder reads the image, a detector conditioned on your text finds every match, and a presence head keeps closely related phrases apart, so a person in a white shirt and a person in a red shirt stay separate answers. The vocabulary carries over too, built from an annotation process covering more than 4 million unique concepts.
What SAM 3.1 adds is Object Multiplex, Meta's name for how it handles many objects at once. Instead of running one tracker per object, it groups objects into fixed-size buckets and processes each bucket jointly through a shared memory, cutting out repeated work. That pays off most on video, where the number of things being followed drives the bill.
Text is not the only prompt it takes. Points, boxes, and example masks all work, for when the thing you want is easier to click than to describe.
The published numbers
Meta's release notes for SAM 3.1 report the model on video benchmarks, measured against SAM 3 with the same code. On video object segmentation, scored with J&F out of 100, it improved on six of seven standard benchmarks: 79.6 against 78.4 on MOSEv1, 92.7 against 92.2 on DAVIS17, 89.2 against 88.5 on LVOSv2, 85.1 against 84.4 on SA-V test, and 62.3 against 60.3 on MOSEv2, the hardest of the set.
On concept segmentation in video, scored with cgF1 out of 100 for whether every instance of a described concept was found, results are mixed rather than uniform, with the clear gain at 52.9 against 50.8 on YT-Temporal-1B. The efficiency claim is separate and large: about seven times faster inference at 128 objects on one data-centre GPU, with no accuracy given up for it.
All of this is vendor-reported, from Meta's own release notes with no third-party leaderboard behind it. It is a fair before-and-after, since the same team measured both checkpoints the same way, but it is not independent verification. Adoption is thinner than the previous checkpoint's, at just under 64,000 Hugging Face downloads last month.
How it compares to the other segmentation models here
These five models are scored on four different scales, so ranking them by number would be dishonest. cgF1 rewards finding every instance of a described concept, J&F rewards mask overlap across frames, average precision rewards detections against a fixed list of everyday classes, and a structure measure rewards one clean cutout. Nothing converts between them.
Picking comes down to what you can say about the target. Reach for SAM 3.1 (Scene) when you can name it in a phrase, especially if you want one merged mask out the other end, and for SAM 3 if you want the previous checkpoint of the same interface. Reach for SAM 2 (Scene) when you cannot name it and want everything separated instead, RF-DETR Seg Medium when the target is an everyday object and speed decides, and BiRefNet (Subject) when there is one subject to cut out.
What to chain it with
Input is an image plus a short phrase. Output is a mask per match with an instance ID, a bounding box and a confidence score, or one merged mask if you ticked the box, so the phrase becomes the selector for whatever comes next.
Two chains use it well. With the merge box ticked, prompt for faces or screens and pass the single mask to Selective Blur, which blurs only those regions and leaves the rest sharp. Without it, prompt for your product, extract the region with Color Pick by Class, and rebuild the shot with Layer Compose over a different background.
Open SAM 3.1 (Scene) in CNAPS Studio, run one image with the merge box off and again with it on, and see which output shape your next step actually wants.
Sources
- The SAM 3.1 release notes in Meta's repository for the release date, Object Multiplex, the video object segmentation and concept segmentation scores against SAM 3, and the seven-times figure.
- github.com/facebookresearch/sam3 for the architecture, the concept coverage figure, the install requirements, and the SAM License with its trade-control terms.
- huggingface.co/facebook/sam3.1 for the checkpoints, the absence of a Transformers integration, and the monthly download count.
- github.com/ZhengPeng7/BiRefNet and github.com/roboflow/rf-detr for the licenses, sizes and metrics of the two models compared above.
- docs.cnaps.ai/ai-model-hub for the CNAPS Studio input, output, merge option and timing.