AI developer MiniMax has released "MiniMax M3," a large language model with roughly 428 billion total parameters, as open weights with the model files included. It is multimodal, handling not only text but also images and video, and on some software-development benchmarks it outscores Google's Gemini 3.1 Pro. The fact that it ships in a form you can run on your own hardware carries real weight for developers.
A Large Model Distributed as Open Weights
MiniMax M3 uses a Mixture-of-Experts (MoE) architecture, in which only the experts needed for a given task are activated. While the total parameter count reaches about 428 billion, only around 23 billion parameters are actually engaged in any single inference pass, a design that balances sheer scale with light runtime cost.
Inputs span text, images, and video, and all three are processed natively without bolting on extra conversion components. The model had already been available via API beforehand, and this release newly opens up the model weights themselves to the public. In other words, a model you could previously only touch through the cloud can now be downloaded into your own environment for testing.
A Custom Attention Mechanism That Lifts Speed
The feature MiniMax M3 puts front and center is its in-house attention mechanism, "MiniMax Sparse Attention (MSA)." Built to handle long context efficiently, it secures a vast context window of up to one million tokens.
The speed gains are pronounced as well. Compared with the previous-generation "MiniMax M2," the company reports roughly a ninefold improvement in the prefill stage that reads input in bulk, and about a fifteenfold gain in the decode stage that writes out text. Furthermore, when processing one million tokens, the per-token compute can be compressed to one-twentieth of the conventional level. The intent is clear: to use architecture itself to tame the long-standing weakness whereby compute costs balloon as context grows longer.
Benchmark Comparison Against Gemini 3.1 Pro
As performance indicators, MiniMax has published comparisons against Gemini 3.1 Pro. On "SWE-Bench Pro," which measures software engineering ability, MiniMax M3 recorded 59.0 percent, surpassing Gemini 3.1 Pro's 54.2 percent. On "MCP Atlas," which gauges the ability to coordinate with external tools, it also came out ahead at 74.2 percent versus Gemini 3.1 Pro's 69.2 percent.
That said, it does not win across the board. On "BrowseComp," which tests autonomous web browsing, MiniMax M3 scored 83.5 percent against Gemini 3.1 Pro's 85.9 percent. On "OSWorld-Verified," which evaluates operating a computer screen, MiniMax M3 came in at 75.2 percent (MiniMax's own posts cite 70.0 percent) versus 76.2 percent for Gemini 3.1 Pro. The picture closest to reality is that it leads in its strong suits while drawing roughly even overall. It is worth keeping in mind that these figures were published by MiniMax itself.
Licensing and System Requirements
The usage terms also matter to developers. Non-commercial use is free of charge, and companies or individuals with annual revenue under 20 million USD (about 3.2 billion yen) may use it commercially simply by notifying MiniMax and including a "Build with MiniMax" attribution.※1 USD = 160 JPY
As for the storage required, the standard bf16 format runs to about 855 gigabytes. However, the heavily quantized 1-bit GGUF format can be squeezed down to roughly 128 gigabytes, and the design deliberately holds the parameter scale in check so it can run on more accessible hardware. For a giant model, leaving room to be tried out on an individual setup is a practical strength.
Summary
MiniMax M3 is an open-weight LLM that releases a 428-billion-parameter large model in a form anyone can download, pairing long-context handling with speed through a custom attention mechanism. It outscores Gemini 3.1 Pro on software-development benchmarks while falling short in areas such as web operation, leaving the two roughly matched in overall capability. As an option you can run on your own GPU, it looks like a strong candidate for developers who want to build AI agents locally.
