The Open Weights
LatestModelsLeaderboardsCompanies
Subscribe
The Open Weights

The daily record of open-source AI. New model releases, leaderboards, and what's coming next — written for people who ship.

Refreshed every 12 hours

Discover

  • Latest releases
  • New today
  • Trending models

Browse

  • All models
  • Companies
  • Categories
  • Leaderboards

About

  • About
  • Editorial policy
  • Hardware estimates
  • RSS feed
  • llms.txt
  • Newsletter

© 2026 The Open Weights. An independent publication.

PrivacyTermsSMSAggregated by Claude · curated by humans.
LatestMiniMaxM3
MiniMaxVision-Language

MiniMax Releases M3, a Multimodal MoE Model

The new open-weight model from MiniMax AI combines vision, coding, and reasoning using a Mixture-of-Experts architecture.

Jun 2, 2026
Major releaseOther
MiniMax-M3

MiniMax AI has entered the open-weight model space with the release of MiniMax-M3, a new multimodal foundation model. Available now on Hugging Face, the model stands out for its sophisticated Mixture-of-Experts (MoE) architecture, a design choice often used to scale models efficiently without a proportional increase in computational cost.

M3 is designed as a versatile system, capable of handling a diverse range of inputs and tasks. Its core strengths lie in its multimodal understanding, allowing it to process and interpret both text and images. Beyond vision, the model is also engineered for coding and complex reasoning, positioning it as a powerful tool for developing more capable AI agents.

The use of an MoE architecture is the key technical detail here. This design activates only relevant neural network pathways—or 'experts'—for a given task. This allows the model to contain a vast number of parameters while keeping inference costs manageable, making powerful capabilities more accessible to a wider range of developers and researchers.

While the model's weights are publicly available, it's released under a custom license rather than a permissive open-source license. This 'open-weight' approach requires users to agree to specific terms for use, reflecting a growing trend where companies release powerful models with some restrictions.

Sources

  • MiniMaxAI/MiniMax-M3

    Hugging Face

    Visit
  • MiniMax M3: The First Open-Weights Model to Combine Three Frontier Capabilities

    Hacker News

    Visit
OlderJD.com Enters Open-Source AI Video with JoyAI-EchoJD · Text → Video · 5 months agoNewerIdeogram 4.0 arrives as an open-weight image modelStability AI · Text → Image · 5 months ago

Get the model

Hugging FaceHacker News

Specs

Active params23B active
Context window1.0M tokens
Size854.2 GB
PrecisionBF16
ArchitectureMiniMaxM3SparseForConditionalGeneration
LicenseOTHER
Downloads129.9K
Likes1.6K

Can you run it?

Needs a multi-GPU server — about 258.2 GB at 4-bit.

  • BF16 (as published)

    Multi-node cluster

    856.2 GB
  • 8-bit

    8×H100 80 GB node

    454.6 GB
  • 4-bit

    4×H100 80 GB · 512 GB Mac Studio

    258.2 GB
Your machine
GGUF builds
  • Mixture-of-experts: all 427B parameters must be resident even though only 23B are active per token — active count affects speed, not memory.
  • KV cache sized for 8,192 tokens of context; longer prompts need proportionally more.

Based on Hugging Face weights metadata. Assumes an 8K context and ~1 GB runtime overhead; actual needs vary. How we estimate


Modalities

Vision-LanguageAny-to-AnyCodeReasoning

The Weekly Weights

Every open release that mattered, one email a week.

0 comments

No comments yet. Be the first to weigh in.

More from MiniMax

All MiniMax releases →
MiniMax H3 Image
MiniMax/Text → Image

MiniMax Releases H3 Image, Its Open Text-to-Image Model

The Chinese AI lab enters the open image-generation space with a single model that both creates and edits pictures from text.

Nov 3, 2026
MiniMax-M3.1-Flash
MiniMax/Vision-LanguageMulti-GPU server

MiniMax Debuts M3.1 Flash, a Multimodal MoE Model

The new release pairs a mixture-of-experts design with agentic, coding, and video-understanding abilities in a single open model.

Oct 17, 2026
MiniMax-Music3
MiniMax/MusicRuns anywhere

MiniMax Opens Its Music3 Text-to-Music Model

The Chinese AI lab brings its music generation system to Hugging Face, expanding the roster of open audio models.

Aug 7, 2026

More in Vision-Language

All Vision-Language →
Mirage
Cactus Compute/Vision-Language

Cactus Compute's Mirage brings vision tasks on-device

The new open vision-language model targets detection, segmentation, and visual Q&A without relying on the cloud.

Nov 5, 2026
Inkling-2-Small
Thinkingmachines/Vision-LanguageMulti-GPU server

Thinking Machines ships Inkling-2-Small, an open MoE VLM

The Apache-2.0 vision-language model uses a mixture-of-experts design and lands openly on Hugging Face.

Nov 4, 2026
LongCat 2.5
Meituan/Text / LLMDatacenter

Meituan Releases LongCat 2.5, an MIT-Licensed MoE

The Chinese delivery giant expands its open-model push with a multimodal, agentic mixture-of-experts system.

Nov 4, 2026