MiniMax just released M3—which many people informally call “MiniMax 3.0”—and on paper it’s impressive. They present it as the first open-weight model with three frontier capabilities at once: coding, a one-million-token context, and native multimodality. Let me be clear about my stance from the start: I have faith that it’s a good model, but it won’t be my daily driver.
What M3 Brings to the Table
The numbers MiniMax publishes are not vapor. M3 runs on a proprietary architecture, MiniMax Sparse Attention (MSA), with a context window of up to 1M tokens (a guaranteed minimum of 512K). The multimodality isn’t a patch: they say they rebuilt the entire data pipeline to train it from step zero.
There are striking figures on agentic benchmarks. On BrowseComp it scores 83.5, surpassing Opus 4.7 (79.3). On their PostTrainBench, where the model autonomously trains other models, it lands third (37.1), behind only Opus 4.7 (42.4) and GPT-5.5 (39.3). And they show powerful demos: reproducing an ICLR paper over 12 hours of autonomous execution, or optimizing a CUDA kernel with a 9.4× speedup after 147 iterations.
As an engineer, that strikes me as genuinely good. An open-weight model fighting in that league is excellent news for everyone.
Where My Reservation Lies
My skepticism isn’t against the benchmarks. It’s against a specific experience that repeats with models that aren’t absolute top-tier: there comes a point where I need more quality and the model simply can’t give it to me.
And here’s the important nuance: that ceiling isn’t solved with scaffolding. I can put subagents to review the code, verification layers, self-critique loops, all the orchestration you want. That improves consistency and reduces silly errors, but it doesn’t raise the reasoning ceiling of the base model. If the model doesn’t “see” the correct solution, a thousand reviewing subagents won’t invent it. They’ll only confirm, with more steps and more cost, the same limitation.
Real Work Exposes the Limits
For bounded tasks—scaffolding, mechanical refactors, generating boilerplate, retrieval over long contexts—M3 will probably perform more than well. There, the 1M context and agentic capabilities are a real advantage.
The problem appears in hard work: the change that touches five systems at once, the subtle bug that demands understanding an entire abstraction, the design decision where the model has to hold a lot of mental state and reason for real. That’s where, in my experience, GPT-5.5 and Claude Opus still make a difference that isn’t about nuance, but about “it solves it or it doesn’t.”
I Don’t Expect It to Cover My Usage
To be fair: I’m not asking M3 to be something it doesn’t claim to be. It’s an open-weight model with an enormous value proposition—open frontier capability, deployable, with a competitive token plan. For a lot of people and a lot of use cases, it’ll be more than enough.
But I don’t expect it to reach the usage I give to a GPT-5.5 or a Claude Opus. My workflow constantly pushes against the upper limit of quality, and there’s no agent architecture that compensates for a lower model ceiling.
My Verdict
M3 gives me faith, and I mean that seriously. It’s a step forward for the open ecosystem and I’ll keep an eye on it. But “good model” and “daily driver for my most demanding work” are two different categories, and for now M3 sits in the first. I’ll use it for what it does well, without asking it for what I know it won’t give me.
