ONLINEAGENT_OPS 2026.Q3 HOME HUBSMODELSBUILDINGCOMPANIESFILMMAKINGSECURITYLIVING WITH ITSKILLSALL HUBS, ONE TERMINAL → ARTICLES CRAFT RECORD BLOG TOOLS SKILLS SEARCH /
HOME›ARTICLES›SAM 3.1: WHAT META CHANGED, AND WHY IT IS TWICE
ARTICLES · EXPLAINERBEGINNER

SAM 3.1: What Meta Changed, and Why It Is Twice as Fast

Meta’s segmentation model now tracks up to sixteen objects in one pass instead of one pass each. The published figures, and where the model already runs.

READ4 min
WORDS865
SECTIONS5
TYPEEXPLAINER
CHECKED1 OCT 26

SAM 3.1 made object tracking twice as fast by stopping the model doing the same work sixteen times. Meta released it on 27 March 2026 as a drop-in replacement for SAM 3, and the whole change is one idea: track many objects in a single pass instead of one pass each.1

TL;DR — THE SHORT VERSION

SAM 3.1 finds the things you name in a picture or video and outlines each one exactly. You can download it after accepting Meta’s conditions, try it online, or call it on Meta’s Model API.

  • Drop-in replacement. Meta’s own word for it — no rearchitecting required.
  • Up to 16 objects in one forward pass, where each object used to need its own.
  • 16 to 32 frames per second on a single H100, on videos with a medium number of objects.
  • Cheaper hardware becomes viable, which is the practical point of the release.
  • The weights are gated. Hugging Face asks you to accept conditions before downloading.

What changed from SAM 3

Only the video path, and only how it handles more than one object. Meta: “Previously, each object required its own dedicated pass, but with multiplexing, SAM 3.1 processes all tracked objects together, eliminating redundant computation and memory bottlenecks.”

The number attached to that: the model can “track up to 16 objects in a single forward pass”, which “doubles the processing speed for videos with a medium number of objects, increasing throughput from 16 to 32 frames per second on a single H100 GPU”.2

WHAT MULTIPLEXING CHANGES
The same objects, tracked two ways. SAM 3.1 stops repeating the pass for each one.
SAM 3 · A PASS PER OBJECTobjectobjectobjectobjectpasspasspasspassSAM 3.1 · ALL OBJECTS, ONE PASSobjectobjectobjectobjectOne forward passup to 16 objects processed together
Meta AI, SAM 3.1, 27 Mar 2026, read at source 16 Sep 2026: “Previously, each object required its own dedicated pass, but with multiplexing, SAM 3.1 processes all tracked objects together, eliminating redundant computation and memory bottlenecks.” Four objects are drawn for legibility; the model can “track up to 16 objects in a single forward pass”.

Meta also claims it helps in crowded scenes rather than just going faster: the approach “streamlines performance and enhances accuracy in crowded scenes”.

TAKEAWAY

This is an efficiency release, not a capability release. What SAM 3.1 can recognise is what SAM 3 could recognise — it just stops paying for the same computation once per object.

What SAM 3 is, if you missed it

A segmentation model you can prompt with words. Meta describes it as “a unified model for detection, segmentation, and tracking of objects in images and video using text, exemplar, and visual prompts”. The model card puts the scale of the vocabulary bluntly: it handles “over 50x more unique concepts than existing benchmarks”.3

In plain terms: you type a short noun phrase and it finds every instance of that thing, in a still or across a clip.

Where you have already seen it

Meta names the products. SAM 3 and SAM 3D “are being used to enable the new View in Room feature on Facebook Marketplace”; new effects are planned for the Edits app, where “Creators can apply dynamic effects to people or objects in their videos — simplifying a complex editing workflow to just one tap”; and Meta says new creation experiences “will be coming to Vibes on the Meta AI app and meta.ai on the web, where people can use AI visual creation tools and remix existing AI-generated videos”.

Two of those three are future tense in Meta’s own sentence. This page has not checked whether the Edits effects or the Vibes features shipped.

Getting hold of it

What Meta published with SAM 3.1, read at source 16 Sep 2026
ArtefactWhere
Model checkpointHugging Face, behind an access gate
Codethe SAM 3 GitHub repository
Research paperlinked from the announcement
Try it without installingSegment Anything Playground
Call it from your own codeMeta Model API, model ID sam-3.1 (listed by 1 Oct 2026)

Meta, facebook/sam3.1, read at source 16 Sep 2026: the page says “Log in or Sign Up to review the conditions and access this model content”, and that “This repository hosts only the SAM 3.1 model checkpoints — there is no Hugging Face Transformers integration.”

That second sentence matters if you assumed a one-line load. Meta points you at the GitHub repository “for installation, code, usage examples, and full documentation”.

There is now a no-download route. Meta’s Model API, a paid web service you call from your own code, lists SAM 3.1 as sam-3.1. You send a short text prompt, such as “dog”, and get back a box and an exact outline (a “mask”) for every match. For costs, see Meta’s Model API page; this site adds no new prices.4

What this page could not verify

  • The speed claim. Meta’s figure, on Meta’s hardware, not reproduced here.
  • The accuracy claim in crowded scenes. Stated in the announcement without a benchmark in the text.
  • Whether the Edits and Vibes features shipped. Announced as coming.
  • What the access conditions say. The gate requires an account this site did not use.
  • When SAM 3.1 was added to the Model API. Meta’s docs page carries no date.
SOURCES
  1. 1Meta AI, SAM 3.1: Faster and More Accessible Real-Time Video Detection and Tracking, 27 Mar 2026, read at source 16 Sep 2026.↑
  2. 2Meta AI, SAM 3.1, 27 Mar 2026, read at source 16 Sep 2026.↑
  3. 3Meta, facebook/sam3.1 model card, read at source 16 Sep 2026SAM 3 “detects, segments, and tracks objects using text or visual prompts such as points, boxes, and masks”. ↑
  4. 4Meta, Meta Model API docs overview, undated, read at source 1 Oct 2026model ID sam-3.1; “SAM 3.1 segments images and video through Model API. Name an object in a short text prompt and get back a box and a pixel-accurate mask for every match, and in video follow each object across frames.” ↑

Meta AI, SAM 3.1: Faster and More Accessible Real-Time Video Detection and Tracking With Multiplexing and Global Reasoning, 27 March 2026 · Meta, facebook/sam3.1 model card. Both read at source on 16 September 2026.

Performance figures are the vendor’s.

Free printable AI cheat sheets, and AI news every week.

Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.

SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK
aipromptgeneer://$ about this siteAI PROMPT GENEERAI explained in plain English. Every claim links to its source, with the date it was checked.NEWSLETTERWhat changed in AI each week, in plain English.SUBSCRIBE FREE ↗AI PROMPT GENEER · MACHINE ASSISTANCE DISCLOSEDEVERY ARTICLE CARRIES ITS OWN CHECKED DATE