Jin Daily AI Trivia: MiniMax M3 Launched!! Good or Bad?

Jin Daily AI Trivia: MiniMax M3 Launched!! Good or Bad?

Open-source AI models are now entering the 1 million token multimodal era, bringing capabilities that until now were only available in a small handful of closed-source models that could combine frontier-level coding, million-token context windows, and multimodal understanding in a single system.

MiniMax M3 comes with a new MSA design, which is based on GQA. The idea is to reduce token memory pressure, allowing the model to run well at 512K tokens and extend up to a 1M context window.

DeepSeek V4 also reached a 1M token context window with its own DSA design, but MiniMax M3 is different because it is trained natively as a multimodal model.

M3 is MiniMax trying to bring that full frontier package into the open world.

If M3 follows what MiniMax M2.7 was already good at, we can expect strong long-context agentic tool-calling usage. Now with multimodal support and a much larger context window, it should get even better.

But here is the catch.

MiniMax has always been a bit sneaky with licensing.

Last time, the M2.7 license stated that commercial usage required written approval from the MiniMax team. I believe M3 may follow the same direction as well, so please don’t assume it is fully free for commercial use until you check the license properly.

What worries me more is that MiniMax seems to have silently dropped its lowest coding/token plan.

The old $9 discounted plan is gone. Now it starts at $20.

With a bigger model and more thinking steps consuming more token usage now, MiniMax M3 no longer feels like the cheapest low-cost AI model for AI agents.

So is MiniMax M3 good?

Technically, yes. And we soon can download the model and deploy locally on our local machine. For open-source multimodal, long-context, coding, and agentic use cases, this is a very serious model.

But is it still the cheap AI agent king? That part is getting harder to say.

Hope you learned something new today!!

See ya!!