The short version
MiniMax M3 is an open-weight AI model that runs frontier-level coding, holds a million tokens of context, and reads images and video, all at once. MiniMax released it on June 1, 2026, and the AI world started talking the same day. The reason is simple. Models that do one of those things well are common. A model you can download and host yourself that aims to do all three is not. If your team is weighing which model to build on, our AI development services team has been tracking releases like this closely, and M3 is worth a careful look.
This post breaks down what M3 is, what it can and cannot do, what it costs, and how it compares to GPT-5.5. We will also be honest about the parts that need more proof.
MiniMax M3 is a large language model with open weights, released by the Shanghai lab MiniMax. “Open weights” means the model files can be downloaded and run on your own hardware, instead of being locked behind a company’s API.
At launch, the weights were not out yet. MiniMax said it would publish them, plus a full technical report, on Hugging Face and GitHub within about ten days. So the model you read about today becomes something you can host yourself soon after. You can read The Decoder’s launch coverage for more on the rollout.
The pitch is that M3 is the first open model to fuse three frontier features in a single design. Most open models pick one strength. M3 wants all three.
The MiniMax M3 features that matter most are its coding skill, its huge context window, its multimodal input, and the architecture that makes the whole thing affordable.
A 1-million-token context window
M3 can take in up to 1,000,000 tokens at once. That is enough to hold a full mid-sized codebase, a long legal file, or hours of chat logs in active memory. Output can run up to 512,000 tokens.
M3 reads text, images, and video as input. MiniMax also showed it operating a desktop computer on its own. That makes it a fit for agent-style tasks, not just chat.
The clever part is the architecture, called MiniMax Sparse Attention. Instead of paying attention to every token, MSA picks only the relevant blocks of context and processes those. A small index branch decides what matters, and the main attention layer handles just that slice.
The payoff is speed and cost. At a million tokens, MSA uses about one-twentieth of the per-token compute of the prior MiniMax generation. MiniMax reports prefill (reading the prompt) runs about 9.7 times faster, and decoding (writing the answer) about 15.6 times faster. Those figures are from MiniMax and have not been checked by outside labs yet.
On its own tests, the headline MiniMax M3 benchmark is 59.0% on SWE-Bench Pro, a hard, real-world coding test. Here is the fuller picture MiniMax shared:
Benchmark | What it measures | M3 score |
SWE-Bench Pro | Real-world software fixes | 59.0% |
Terminal-Bench 2.1 | Command-line agent tasks | 66.0% |
SWE-fficiency | Efficient code changes | 34.8% |
KernelBench Hard | Low-level GPU kernels | 28.8% |
BrowseComp | Autonomous web browsing | 83.5 |
Now the caveats, and they are real. These scores come from MiniMax, run on its own setup. Independent results from third parties were not out at launch. MiniMax also compared M3 against Claude Opus 4.7, even though Anthropic had already shipped Opus 4.8 a week earlier. Against the newer Opus 4.8, the coding gap is wider. Tech Times covered this gap in detail.
Plain advice: treat launch benchmarks as a hint, not a verdict. Build a small test set from your own real tasks and run M3 against it before you commit.On the question of MiniMax M3 vs GPT-5.5, the answer depends on what you value. M3 wins on price, on context size, and on the fact that you can host it yourself. GPT-5.5 is a closed, polished product with broad, proven results.
On SWE-Bench Pro, M3’s 59.0% edges past GPT-5.5’s reported 58.6%. That is close, and it is one test. M3’s bigger advantages are the 1-million-token window and a much lower price. The catch is maturity. GPT-5.5 has a long track record and wide tooling support. M3 is days old, and some of its claims still need outside proof.
If your workload is heavy on coding or long documents and you watch costs, M3 deserves a trial. If you need a settled, well-supported model for mixed work, GPT-5.5 is the safer default for now. A short discovery session, like the ones our innovation consulting team runs, can help you score both against your own goals before you pick.
The MiniMax M3 pricing is the part that turned heads. Standard rates are $0.60 per million input tokens and $2.40 per million output tokens.
For the first week after launch, MiniMax cut that in half: $0.30 input and $1.20 output per million tokens. That puts M3 at roughly one-fifteenth to one-twenty-fifth the operating cost of closed frontier models. VentureBeat reported M3 reaching that performance tier at 5 to 10 percent of the usual cost.
There are also monthly token plans, around $20, $50, and $120. And once the open weights ship, self-hosting removes the per-token bill entirely, though you then pay for your own hardware.
Here is why the price matters in practice. A feature that costs over a cent per call on a premium model can cost a fraction of that on M3. Cheap features that were not worth shipping suddenly pencil out.
Open weights change the data story. When you self-host M3, your prompts and your customers’ data stay on your own infrastructure. Nothing has to travel to a third-party API. For teams handling regulated data, that is a big deal.
This helps with rules like GDPR, where data residency and control matter. Healthcare and finance teams who must keep records in-house gain a real option. Building that kind of setup safely is its own project, and it is the sort of work our AI development company handles end to end.
Two cautions, though. First, the hosted MiniMax API is run by a company based in China. China’s 2017 National Intelligence Law can require local firms to cooperate with the government. For sensitive data on the hosted API, that belongs in your risk review. Self-hosting the open weights sidesteps most of that concern.
Second, check the license before you build a commercial product on M3. An earlier MiniMax model restricted commercial use without written permission. At launch, M3’s final license terms were not confirmed. Read the actual license when the weights drop, not the headline.
