SAM 3 and SAM 3.1 are two releases of the same model: Meta's Segment Anything, which detects, segments and tracks objects in images and video from a click, a box or a short description. SAM 3.1 is not a rewrite. Meta describes it as a drop-in replacement whose one addition is object multiplexing, and the practical question is whether that difference reaches you. For most single-subject work it does not; for anything with several subjects at once it is the whole point.
SAM 3.1 vs SAM 3: the specs
Meta's published figures for each release.
| SAM 3.1 | SAM 3 | |
|---|---|---|
| Released | March 27, 2026 | Announced with the SA-Co benchmark |
| Relationship | Drop-in replacement for SAM 3 | The base release |
| How many objects per pass | Up to 16, multiplexed | One object per pass |
| Video throughput | 32 frames per second on one H100 | 16 frames per second on one H100 |
| Cost of extra objects | Shared compute across tracked objects | Scales linearly, one masklet each |
| Prompting | Click, box, mask, text or image exemplar | Click, box, mask, text or image exemplar |
| Images and video | Both | Both |
| Single-image speed | Unchanged from SAM 3 | About 30 ms with 100+ objects, on an H200 |
| Weights | Checkpoint published | Weights, code and evaluation set published |
| On Morphic | Yes, as subject tracking and cutouts | No |
What SAM 3.1 changes
- Many objects stop competing for the machine. SAM 3 tracked each object with its own pass, sharing only the per-frame picture it had already encoded, so a second subject cost roughly a second run. SAM 3.1 processes the tracked objects together in one forward pass, up to sixteen of them.
- Throughput doubles on Meta's own measurement. Meta puts the gain at 16 to 32 frames per second on a single H100 GPU for video with a medium number of objects, and attributes it to the redundant work multiplexing removes.
- Crowded shots become practical. The reason to care is not raw speed, it is the shots you would previously have avoided: a group, a product shelf, traffic, a stage with several performers.
What SAM 3 already did, and 3.1 keeps
- Promptable concept segmentation. A short noun phrase such as "the cyclist" finds every instance of that concept, rather than only the one thing you clicked. Meta reports roughly double the cgF1 score of existing models on its SA-Co benchmark, and says people preferred its results to OWLv2 at about three to one.
- The same four ways to prompt. Clicks, boxes and masks carry over from the earlier Segment Anything models, and text and image exemplars arrived with SAM 3. Nothing about this changed in 3.1, so anything written for SAM 3 still applies.
- The same limits. Long compositional phrases are still out of scope, and Meta still describes the model as weak on fine-grained, out-of-domain concepts until it is fine-tuned on them.
Which one you are using on Morphic
Morphic runs SAM 3.1. In Compose, Morphic's timeline editor, it is the Track subject mask: mark a subject on one frame, pick a direction, and get back a mask video that follows it. On the Canvas it is Auto-segment and Separate as layer for stills. There is no version switch to make, and nothing written for SAM 3 needs rewriting, because the prompting is identical.
SAM 3.1 on Morphic: the picture on the left, the mask it returned on the right.
Where the 3.1 difference lands in practice is the clip with several things to isolate. Tracking four people in a scene is four marked subjects and four masks either way; what changed is how much machine time sits behind them.
For the full feature set, see the SAM 3.1 model page. For the walkthrough, read how to track a subject with SAM 3.1, and for the job itself, AI rotoscoping.
FAQs
Meta presents 3.1 as a speed and capacity change rather than an accuracy one. The documented difference is object multiplexing and the throughput it buys, not a new quality claim.
No. The ways of prompting are the same: a click, a box, a mask, a short text phrase or an image exemplar. Meta calls 3.1 a drop-in replacement.
Meta says up to 16 objects in a single forward pass. On Morphic a clip also holds up to sixteen masks, one per tracked subject.
No. Morphic runs SAM 3.1 only, and since the prompting is unchanged there is nothing to migrate.
No. That limit is Morphic's, not the model's: a clip holds up to sixteen masks, tracked or drawn, in any combination. Meta's sixteen-objects-per-pass figure is about how the model processes them, not about how many you are allowed.
From Meta's own post about the model. The 16 to 32 frames per second figure is measured on a single H100 GPU; the roughly 30 millisecond single-image figure for SAM 3 is on an H200 with more than 100 objects detected.