NEWMVS for embeddings. Managed for files. Both on object storage.Vectors or files. Pick a path.Start →

Start here

Vector Store (MVS)

Bring your own vectors. Dense, sparse, and BM25 search on object storage.

Managed Indexing

Connect a bucket and auto-extract scenes, faces, OCR, transcripts, and embeddings.

Build

Compose multi-stage search in <100ms: filter, join, rerank.

Feature Extractors

Typed pipelines for faces, scenes, transcripts, OCR, fingerprints.

S3, GCS, R2, Mux, LangChain, MCP, and more. Connect your stack.

Generate and store embeddings from 50+ models, then search them.

By Industry

Map, search, and reuse the moments that perform. Plugs into iconik & Mux.

Talent search, brand safety, creative analytics.

Scene search, recommendation, archive access.

Visual search, PDP enrichment, catalog QA.

Lecture search, transcript Q&A, content safety.

View all solutions →

By Use Case

Face & Person Search

Find anyone across video libraries in milliseconds.

IP & Copyright Detection

Logos, songs, faces: one pipeline, one report.

Visual Taste & Recs

Scene-similarity ranked recommendations with RL.

Brand & Ad Safety

Pre-publish content screening at bid-time speeds.

View all use cases →

Build

API reference, SDKs, recipes, and architecture guides.

Launches, deep dives, and field notes from our engineers.

Browse supported HuggingFace models by task and modality.

See what teams are building with Mixpeek.

Education

Vendor-neutral deep dives on perception, retrieval, and embeddings.

Best-of comparisons: vector DBs, embedding models, moderation APIs.

Multimodal University

Fundamentals of multimodal retrieval, modules + certs.

Every term you need: embeddings to re-rankers.

Talks, demos, and customer sessions on demand.

Visual explainers: embeddings, chunking, hybrid search, reranking, RAG.

Papers behind multimodal search — MUVERA, SAM 3, and more — explained.

Mixpeek vs. Pinecone, Weaviate, Twelve Labs, more.

Mission, team, and the multimodal vision.

We're hiring across research, infra, and design.

Talk to sales, support, or press.

45-min working session on your data — leave with a running notebook.

Sign in Request Demo Get started →

Meta's SAM 3: Type a Phrase, Segment Every Instance

81

Retrieval & Search Research

Mixpeek Team

July 20, 2026

Summary

SAM 1 and 2 returned one mask per click. SAM 3 (Meta, November 2025, open weights) introduces promptable concept segmentation: a short noun phrase or exemplar image makes one model detect, segment, and track every instance of that concept across images and video. A shared Perception Encoder feeds a DETR-based detector and a SAM 2 memory tracker, trained on 4M+ auto-annotated concepts. Results: 55.7 cgF1 on SA-Co Gold vs 24.5 for OWLv2, 75-80% of human performance, 30ms per image with 100+ objects; SAM 3.1 tracks 16 objects in one pass at 32 fps. Already powers Instagram Edits and Marketplace View in Room. Explore: ai.meta.com/blog/segment-anything-model-3

sam3segmentationcomputer-visionmetavideo-understandingobject-tracking

About this video

SAM 1 and 2 returned one mask per click. SAM 3 (Meta, November 2025, open weights) introduces promptable concept segmentation: a short noun phrase or exemplar image makes one model detect, segment, and track every instance of that concept across images and video. A shared Perception Encoder feeds a DETR-based detector and a SAM 2 memory tracker, trained on 4M+ auto-annotated concepts. Results: 55.7 cgF1 on SA-Co Gold vs 24.5 for OWLv2, 75-80% of human performance, 30ms per image with 100+ objects; SAM 3.1 tracks 16 objects in one pass at 32 fps. Already powers Instagram Edits and Marketplace View in Room. Explore: ai.meta.com/blog/segment-anything-model-3

Frequently Asked Questions