MiniMax M3

MiniMax's 428B multimodal mixture-of-experts, 23B active: sparse attention for 1M context, image and video input, and a community license.

In short. MiniMax M3 is a large AI model for coding and agent work that can also read images and video. Its new attention method keeps very long inputs, up to a million tokens, affordable. It is a server-class model, and its license lets you use it freely for non-commercial work but sets conditions for products.

Summary

MiniMax released M3 on 1 June 2026. It is a native multimodal mixture-of-experts model with about 428 billion parameters, of which about 23 billion are active per token. It reads text, images and video, writes text, and has a 1-million-token context window. According to the model's configuration it has 60 layers and 128 experts, with 4 routed experts and 1 shared expert used per token. MiniMax also says it can operate a desktop computer, and the model supports three reasoning modes through a thinking setting: enabled, adaptive and disabled.

The architectural change is MiniMax Sparse Attention (MSA). Instead of letting every token attend to the whole cache, MSA scores blocks of 128 tokens, picks the best blocks for each query and runs attention only over those; vLLM describes the design as a hybrid of dense and sparse attention. MiniMax says that at 1 million tokens this cuts per-token compute to one twentieth of its previous generation, with more than 9 times faster prefill and 15 times faster decoding. These numbers are its own and have not been independently confirmed. In its launch post MiniMax reports 59.0% on SWE-Bench Pro, 66.0% on Terminal-Bench 2.1 and 74.2% on MCP Atlas, measured on its own infrastructure, with Claude Code as the scaffold for the SWE-Bench tests.

The BF16 checkpoint is about 854 GB, and an MXFP8 version is about 444 GB, so even the smaller one needs a multi-GPU server. vLLM documents day-0 serving support in a June 2026 blog post. The weights use the MiniMax Community License: free for non-commercial use, and for commercial use you must display "Built with MiniMax M3" and send a one-time notice, or get written authorization if the product earns more than US$20 million a year. MiniMax M3.1 Flash, released in October 2026, keeps the core design of M3 in a version trained for 4-bit weights, at about 252 GB.

Good for

  • Long-context coding agents that read whole repositories, logs and documents in one session
  • Tasks that mix code with screenshots, documents or video, and agents that operate a desktop
  • Teams with multi-GPU servers that can meet the license conditions
  • Choosing a reasoning mode per request, from no reasoning for speed to always-on reasoning

Watch out for

  • The BF16 download is about 854 GB, and the MXFP8 download is about 444 GB, so this is not a laptop or single-GPU model
  • The license is not Apache 2.0 or MIT; commercial use needs the "Built with MiniMax M3" credit and a notice email, and revenue above US$20 million needs written authorization
  • Speed-up figures and benchmark scores come from MiniMax, and its coding tests ran on its own infrastructure
  • MiniMax M3.1 Flash is newer and much smaller to download, so compare the two before you deploy

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Estimates for the M3 (BF16) variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.

Variants Parameters BF16 FP8 / INT8 4-bit (Q4) Fits on
M3 (BF16) BF16 428B 23B active 984 GB 492 GB 295 GB Big-memory desktop
M3 MXFP8 FP8 428B 23B active 984 GB 492 GB 295 GB Big-memory desktop

Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software.

VRAM calculator

The license in plain words

A license that grants rights for non-commercial purposes and then lays out the steps for commercial use: show a credit line, send a notice, and get written authorization if your product earns over US$20 million a year. Military use is banned. Get a legal read before shipping a product on it.

  • Attribution or model name must be shown
  • Free commercial use only below a revenue cap
  • Acceptable-use policy applies

More about MiniMax Community

See also

Frequently asked questions

What is MiniMax M3?
MiniMax M3 is a large AI model for coding and agent work that can also read images and video. Its new attention method keeps very long inputs, up to a million tokens, affordable. It is a server-class model, and its license lets you use it freely for non-commercial work but sets conditions for products.
Can I use MiniMax M3 commercially?
Yes, with conditions. MiniMax M3 is released under the MiniMax Community, which allows commercial use but adds extra rules: Attribution or model name must be shown; Free commercial use only below a revenue cap; Acceptable-use policy applies. Read the license text before you ship a product.
Is MiniMax M3 open source?
MiniMax M3 is open-weight, not open source in the strict sense. You can download the model files, but the MiniMax Community adds conditions that the Open Source Initiative definition does not allow.
Can I run MiniMax M3 on my own computer?
Only on a machine with a lot of memory. The smallest variant, M3 (BF16), needs about 295 GB at 4-bit precision. Desktops with 128 to 512 GB of unified memory can hold it, at modest speed.
When was MiniMax M3 released?
MiniMax M3 was released by MiniMax on June 1, 2026.
How long a context window does MiniMax M3 have?
MiniMax M3 supports a context window of up to 1M tokens.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
MiniMax M3.1 Flash MiniMax 428B 23B active 1M MiniMax Community (Conditional) 277 GB Big-memory desktop
DeepSeek V4 DeepSeek 284B – 1.6T 49B active 1M MIT (Permissive) 176 GB Big-memory desktop
GLM-5.3 Z.ai (Zhipu AI) 753B 40B active 1M GLM-5.3 License (Conditional) 520 GB GPU server
Kimi K3.1 Pro Moonshot AI 2.8T 104B active 1M Kimi K3 License (Conditional) 1.72 TB Multi-node cluster