MiniMax M3 is an open-weight frontier model released on June 1, 2026 that pairs a 1-million-token context window with frontier-level coding performance and pricing far below the closed leaders. In plain terms: it is a model you can download and run yourself, it can hold roughly a mid-sized codebase in a single prompt, it scored about 59% on SWE-bench Pro (the agentic coding benchmark), and it costs a fraction of what GPT-5.5 or Claude Opus charge. That combination is the news. Open weights and low price are common at the small end of the market. Open weights plus a genuinely competitive frontier result is not.
This post covers what MiniMax M3 is, how its sparse-attention architecture makes a 1M context affordable, what the coding numbers actually say, how the pricing compares, and who should care. If you build products on AI, run your own infrastructure, or just want to understand why a Chinese lab’s open release is a talking point this summer, read on.
Who makes MiniMax, and what "open-weight" means here
MiniMax is a Shanghai-based AI company. It is not a household name in the West the way OpenAI or Anthropic are, but it has shipped a steady line of capable models, and M3 is its most ambitious release to date. The label that matters most is open-weight.
Open-weight means the trained model parameters (the weights) are published for download. You can run the model on your own hardware, fine-tune it, and inspect it. That is different from a closed model like GPT-5.5 or Claude, where you only ever touch the model through an API and never see the weights.
Open-weight is also not the same as fully open-source. A truly open-source model would ship under a permissive license (MIT, Apache) with no restrictions on commercial use. MiniMax M3 is released under a community license that is more restrictive than that. Reporting at launch described it as similar in spirit to a modified-MIT license, but the exact terms should be read before you build a commercial product on it. Flag for anyone planning production use: check the current license text on the model’s repository before you commit. For a fuller treatment of where models sit on the open-to-closed spectrum, our explainer on Llama, Meta’s open-weight family, walks through the same distinction in detail.
If the term "frontier model" is doing heavy lifting for you, we define it precisely in our piece on what a frontier model is. The short version: a frontier model is one at or near the leading edge of capability. Until recently, every frontier model was closed. M3 is one of the reasons that is no longer true.
The 1-million-token context window
M3 accepts up to 1,048,576 tokens of input, the round number everyone shortens to "1M context." Tokens are the chunks of text a model reads and writes; if you want the mechanics, our primer on how AI models break text into tokens covers it. As a rough rule, 1 million tokens is somewhere around 700,000 to 750,000 words of English, or a mid-sized code repository, or a stack of long PDFs.
Why does that matter? Context is the model’s working memory for a single request. A bigger context means you can hand the model more material at once without splitting it up or building a retrieval system to feed it in pieces. For coding agents, a 1M window can hold enough of a project that the model reasons about the whole thing rather than one file at a time. For document work, it can read an entire contract set or research corpus in one pass.
M3 also supports a very large output ceiling (reported around 512,000 tokens), which matters for tasks that generate a lot at once, like large refactors or long structured documents.
Sparse attention (MSA) in plain terms
Here is the engineering problem a 1M context creates, and how MiniMax gets around it.
Standard transformer attention compares every token in the prompt against every other token. That is the mechanism that lets a model relate a word at the end of a document to one at the beginning. The catch is that the cost grows with the square of the length. Double the input and you roughly quadruple the attention work. At a million tokens, that quadratic cost is brutal: it makes long contexts slow and expensive to run, which is why most models that advertise huge windows are painful to actually use at full length.
MiniMax’s answer is MiniMax Sparse Attention (MSA). "Sparse" means the model does not compare every token to every other token. Instead it attends selectively, focusing compute on the token relationships that matter and skipping the ones that do not. The result is that cost grows far more gently as the context gets longer.
The numbers MiniMax reports for MSA at million-token length are the selling point:
- Roughly one-twentieth the per-token compute at 1M context compared to its previous generation.
- Around 9x faster prefill, prefill being the phase where the model reads and processes your input.
- Around 15x faster decoding, decoding being the phase where it generates the output.
Those are the vendor’s own figures, measured against MiniMax’s prior model rather than against competitors, so read them as "how much MSA improved on their last release" rather than an independent ranking. The mechanism is real and the direction is right: sparse attention is the practical reason a 1M window can be priced cheaply instead of being a demo feature you never turn on.
The coding results
M3’s headline benchmark is SWE-bench Pro, where it scored about 59%. SWE-bench Pro is a hard agentic coding test: the model is given real software issues from real repositories and has to produce patches that pass the projects’ own tests. It rewards a model that can navigate a codebase, reason across files, and make working changes, not just autocomplete a function.
A 59% result puts M3 in genuinely competitive territory. At launch MiniMax positioned it as on par with GPT-5.5 on that benchmark and approaching Anthropic’s Claude Opus 4.7, which has been the reference point for agentic coding. If you want the closed-model context, we compared the two most recent Opus releases in our Claude Opus 4.7 vs 4.8 breakdown.
A few honest caveats before you treat that number as gospel:
- Benchmarks are directional, not decisive. A single percentage on one test does not tell you how a model behaves on your stack, your languages, or your codebase conventions.
- Vendor-reported scores invite scrutiny. Independent replication tends to lag launch by weeks. The figure here is what MiniMax and early coverage reported; expect third-party numbers to firm it up.
- Agentic coding is more than a score. Tool use, error recovery, and how gracefully a model handles being wrong all matter, and none of that shows up in one benchmark line.
M3 also posts strong results on other agentic and multimodal tests (browsing, computer-use, and video understanding among them), and it is natively multimodal, meaning it takes image and video input alongside text. For most readers evaluating it, the coding result is the one that decides whether it earns a place in the toolchain.
The pricing, and why it is the real story
Capability gets the headlines. Price is what changes behavior. M3’s API pricing at launch is aggressive:
| Model | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|
| MiniMax M3 (standard) | ~$0.60 | ~$2.40 |
| MiniMax M3 (launch promo) | ~$0.30 | ~$1.20 |
| Closed frontier leaders | Several dollars | $10 and up |
The structure matters as much as the headline number. AI pricing is split into input and output, and output almost always costs more, because generating tokens is the expensive half. We unpack why in input vs output tokens. M3’s roughly 4x output-to-input ratio is normal; what is not normal is that both numbers sit an order of magnitude below the closed leaders while the coding score stays in the same conversation.
Now layer the open weights on top. If your workload is large and steady, you are not limited to the API at all. You can self-host and pay for GPUs instead of per-token fees, which flips the economics entirely at volume. That is the combination that makes M3 interesting: cheap if you rent it, and yours to run if you own the hardware.
Who MiniMax M3 is actually for
Not everyone needs this. Here is the honest cut.
- Self-hosters and privacy-sensitive teams. If data cannot leave your environment (healthcare, legal, regulated finance), an open-weight model you run in-house is the whole point. M3 gives you a frontier-adjacent option to run behind your own firewall. Our overview of running local AI models covers the tradeoffs of going self-hosted.
- Cost-sensitive builders at volume. If you are shipping an AI feature and per-token costs are eating your margin, an order-of-magnitude cheaper model that still codes well is worth a serious evaluation.
- Long-context workloads. Whole-repo coding agents, large-document analysis, and anything where splitting the input into chunks is a headache benefit directly from the 1M window plus cheap tokens.
And who should probably wait:
- Small teams with light usage. If you make a few thousand API calls a month, the price difference is noise and the closed leaders’ polish, tooling, and support are worth more than the savings.
- Anyone who needs the absolute top of a specific benchmark. M3 is competitive, not clearly ahead. If your use case lives or dies on the single best model for a narrow task, keep testing the closed frontier too.
- Teams without infrastructure appetite. Self-hosting M3 at full context is a real hardware commitment (multi-GPU, hundreds of gigabytes of memory). The API sidesteps that, but if the license is your reason for choosing it, you inherit the ops burden.
The broader pattern is what to watch. For most of the modern AI era, the frontier was closed and the open models trailed by a clear margin. Releases like DeepSeek’s, which we covered in our look at DeepSeek V4 Pro, and now MiniMax M3 keep narrowing that gap. You do not have to adopt M3 to benefit from it. Competition at the frontier pushes prices down and options up across the board, and that is good for anyone building on this technology.
Frequently asked questions
Is MiniMax M3 free to use?
The weights are published for download, so you can run the model on your own hardware without paying MiniMax per request. That is not the same as “free,” because you pay for the GPUs and operations. Using it through MiniMax’s API costs money per token, though the rates are very low compared to closed frontier models. And “open weight” is not fully open source: the license carries restrictions, so read the current terms before commercial use.
What is a 1-million-token context window in practical terms?
Roughly 700,000 to 750,000 words of English, or a mid-sized code repository, held in a single request. It lets the model reason over a large body of material at once instead of you splitting it into pieces. A large window is a ceiling on how much the model can take in, not a guarantee it reasons perfectly over all of it, so test on your own content.
What is MiniMax Sparse Attention (MSA)?
It is the architecture that makes the 1M context affordable. Standard attention compares every token to every other token, and that cost grows with the square of the length, which makes long contexts slow and expensive. Sparse attention attends selectively instead, focusing compute where it matters. MiniMax reports it cuts per-token compute to about one-twentieth of its prior generation at 1M context, with roughly 9x faster input processing and 15x faster output generation.
How does MiniMax M3 compare to GPT-5.5 and Claude Opus?
On the SWE-bench Pro coding benchmark, M3 scored about 59%, which MiniMax positions as on par with GPT-5.5 and approaching Claude Opus 4.7. It is competitive at the frontier, not clearly ahead. What separates it is being open-weight and dramatically cheaper. For narrow tasks where you need the single best model, keep testing the closed leaders alongside it.
Can I run MiniMax M3 on my own hardware?
Yes, that is the point of an open-weight release. Running it at full capability is a serious commitment, though: reports at launch put full deployment at hundreds of gigabytes of GPU memory, realistically a multi-GPU cluster. Quantization can lower those requirements. If you do not have that hardware or appetite, the API gives you the same model without the ops burden.
