[Models](/models)/[Captioning](/models?category=captioning)/MiniMaxAI/MiniMax-M3

HFScene CaptioningMiniMax Community License

# MiniMax-M3

by MiniMaxAI

Agent-native MoE vision-language model with native video understanding at 1M context

200Kdl/month

428B (23B active)params

[HuggingFace](https://huggingface.co/MiniMaxAI/MiniMax-M3)[Run it on your own data, free](https://studio.mixpeek.com/namespaces/create?mode=managed&utm_source=homepage&utm_medium=model_page&utm_campaign=mvs_onramp)

Identifiers

Model ID

`MiniMaxAI/MiniMax-M3`

## Deploy MiniMax-M3

Single-tenant

Mixpeek has no managed extractor for this model. On a single-tenant deployment you upload the weights and a custom plugin serves them next to the rest of your pipeline.

[Read the customer-hosted guide](https://mixpeek.com/docs/customer-hosted/overview) [Talk to us about single-tenant](/demo)

## Overview

MiniMax-M3 is a sparse mixture-of-experts vision-language model, about 428B total parameters with roughly 23B active per token, trained natively on text, images, and video from the start rather than bolting vision onto a text LLM. Its headline trick is MiniMax Sparse Attention (MSA), which cuts per-token attention compute to about 1/20 of dense attention and delivers 9x prefill and 15x decode speedups at a 1M-token context, so it can reason over long videos and multi-document sessions in one pass.

On Mixpeek, MiniMax-M3 is a strong scene-understanding extractor for video and image collections: it produces grounded descriptions, answers questions about frames, and drives agentic pipelines where an agent inspects footage, decides what matters, and stores the result as searchable metadata. Its long context makes it a fit for whole-clip understanding rather than isolated frames.

## Architecture

Sparse MoE transformer, ~428B total / ~23B active parameters, natively multimodal (text, image, video). MiniMax Sparse Attention (MSA) reduces attention compute and memory so the model sustains a 1M-token context with large prefill/decode speedups over dense attention. Custom modeling code; served via Transformers.

## Mixpeek SDK Integration

IngestRetrieve

// No extractor parameter takes a Hugging Face model id (checked against
// GET /v1/discovery/extractors, which returns 13), so MiniMax-M3 runs
// on your side and the output is upserted through POST
// /v1/namespaces/{namespace\_id}/documents/upsert. On Enterprise the other
// path is to upload the weights instead: POST /v1/namespaces/{id}/models
// accepts the huggingface format and a custom plugin loads them.
const res = await fetch(
  "https://api.mixpeek.com/v1/namespaces/ns\_your\_namespace/documents/upsert",
  {
    method: "POST",
    headers: {
      Authorization: "Bearer API\_KEY",
      "Content-Type": "application/json",
    },
    body: JSON.stringify({
      collection\_id: "col\_your\_collection",
      documents: \[
        {
          document\_id: "asset-00412",
          // The model produces text, so it lands in payload. Give the
          // collection a text vector index and embed that text to make it
          // searchable rather than only filterable.
          payload: { extracted\_text: modelOutput, source\_key: "archive/2026/asset-00412" },
          vectors: { "multimodal-embedding": embeddingOfModelOutput },
        },
      \],
    }),
  },
);

// Managed alternative, if this exact model is not the requirement:
// universal\_extractor@v1 runs google/gemini-embedding-2
// (3072-d) over a bucket, with no inference of your own.

## Capabilities

*   Native video understanding (85.4% on Video-MME-v2)
*   1M-token context for whole-clip and multi-document reasoning
*   Long-horizon agentic and tool-use tasks
*   Grounded image and frame question answering

## Use Cases on Mixpeek

Whole-clip video understanding and captioning for search metadata

Agent inspection of footage where the query decides what to extract

Long-context reasoning over multi-page or multi-clip evidence

Scene descriptions that feed downstream embedding and retrieval

## Common Pipeline Companions

[

Qwen/Qwen3-VL-Embedding-2B

Embed frames + descriptions for hybrid visual-text retrieval

](/model/Qwen/Qwen3-VL-Embedding-2B)[

pyannote/speaker-diarization-3.1

Add who-spoke-when so agents search by speaker and scene together

](/model/pyannote/speaker-diarization-3.1)

## Explore on Mixpeek

[

More Captioning models

Compare alternatives in this category

](/models?category=captioning)[

Best Document AI Platforms

Hand-picked tools & platforms compared

](/curated-lists/best-document-ai-platforms)[

OCR & Document AI Internals

Deep-dive technical guide

](/guides/ocr-document-ai-internals)[

Feature Extractors

See how Mixpeek runs models as extractors

](/extractors)[

Mixpeek Vector Store

Store & search embeddings at scale

](/mvs)[

Pricing

Usage-based pricing for pipelines

](/pricing)[

All Curated Lists

Compare models, APIs & infrastructure

](/curated-lists)

### Specification

FrameworkHF

OrganizationMiniMaxAI

FeatureScene Captioning

Outputtext

Modalitiesvideo, image

RetrieverSemantic Search

Parameters428B (23B active)

LicenseMiniMax Community License

Downloads/mo200K

[

### Research Paper

MiniMax-M3

arxiv.org](https://huggingface.co/MiniMaxAI/MiniMax-M3)

### Build a pipeline with MiniMax-M3

Add this model to a processing pipeline alongside other extractors. Combine with retrieval stages for end-to-end search.

[Run it on your own data, free](https://studio.mixpeek.com/namespaces/create?mode=managed&utm_source=homepage&utm_medium=model_page&utm_campaign=mvs_onramp)

### Alternative Models

[

Salesforce/blip2-opt-2.7b

Scene Captioning

](/model/Salesforce/blip2-opt-2.7b)[

microsoft/Florence-2-large

Scene Captioning

](/model/microsoft/Florence-2-large)[

google/paligemma2-3b-mix-448

Scene Captioning

](/model/google/paligemma2-3b-mix-448)[

Qwen/Qwen3-VL-8B-Instruct

Scene Captioning

](/model/Qwen/Qwen3-VL-8B-Instruct)