MiniMax M3 is an open-weight large language model from the Chinese AI lab MiniMax, released at the end of May 2026 with a context window of roughly one million tokens. It is built for agentic coding and long-document work, it ships under a permissive but not fully open-source license, and it undercuts most Western frontier models on price by a wide margin. If you have seen MiniMax M3 mentioned alongside DeepSeek, Kimi, and GLM and want to know what the fuss is about, this is the decoder.
The pitch from MiniMax is that M3 is the first open-weight release to combine frontier-grade coding, a one-million-token context window, and native multimodal input in a single model. Whether it clears that bar depends on which benchmark you trust, and on who ran it. Here is what is verifiable, what is a vendor claim, and where the model actually fits.
A frontier model you can download
Most of the models a business operator hears about (GPT-5.5, Gemini 3.1 Pro, Claude Opus 4.8) are closed. You reach them through an API, you never see the weights, and you cannot run them on your own hardware. MiniMax M3 is the opposite. MiniMax published the trained model weights to Hugging Face in early June 2026, so anyone with sufficient hardware can download and run it.
That places M3 in the same category as Meta’s Llama, DeepSeek’s V-series, and Moonshot’s Kimi: capable models whose parameters are public. If the phrase is new to you, our explainer on what a frontier model is covers the capability tier these releases are chasing. The novelty here is that MiniMax is claiming frontier-class results while giving the weights away, a combination that was rare even a year ago.
What MiniMax M3 actually is
Under the hood, MiniMax M3 is a Mixture-of-Experts (MoE) model. In an MoE design, the network is split into many specialized subnetworks ("experts"), and only a small fraction activate for any given token. That keeps the compute cost per token far below what the total parameter count would suggest.
Exact numbers are where reporting gets messy. Independent trackers list M3’s total size anywhere from roughly 230 billion to 430 billion parameters, with an active set of only about 10 billion to 23 billion parameters per token. MiniMax has not published a single definitive figure, so treat any specific parameter count you see as an estimate rather than a spec.
The headline architectural change is what MiniMax calls MiniMax Sparse Attention (MSA). Standard transformer attention gets more expensive as the context grows, which is why long-context models have historically been slow and costly. MSA replaces full attention with a block-selection scheme that only attends to the most relevant chunks of the context. MiniMax reports that this cuts per-token compute to roughly one-twentieth of its previous generation at long context, with more than nine times faster prefill (processing your input) and more than fifteen times faster decoding (generating output). Those figures are the vendor’s own, measured against the earlier MiniMax M2 rather than against competitors, so read them as a generational speed claim, not a benchmark result.
M3 is also natively multimodal: MiniMax says it accepts image and video input alongside text, rather than bolting vision on as a separate model.
The 1M-token context window, and why it matters
The specification drawing the most attention is the context window: about 1,048,576 tokens, with an output ceiling reported around 512,000 tokens. A token is a chunk of text a bit shorter than a word, so a million tokens is on the order of 700,000 to 750,000 words. That is enough to hold a large codebase, a full contract set, or a book-length corpus in a single prompt.
Context length is only useful if the model stays coherent across it and if you can afford to fill it, which is where MSA and the pricing come in. It is also why the input and output sides of the bill matter so much at this scale; our piece on input versus output tokens explains why a long context can quietly dominate your costs. A million-token window that costs a fortune to use is a spec-sheet number. MiniMax is pairing the window with aggressive pricing precisely to make it practical.
What the benchmarks say, and who ran them
MiniMax’s headline claim is a score of 59.0% on SWE-Bench Pro, a software-engineering benchmark, which the company says edges out both GPT-5.5 and Gemini 3.1 Pro. That would be a striking result for an open-weight model.
Two caveats keep it honest. First, several of the launch benchmarks were run on MiniMax’s own infrastructure with its own agent scaffolding, so they are vendor-reported until independent labs reproduce them. Second, benchmark leads at this level are narrow and shift from month to month. For an independent read, the third-party evaluator Artificial Analysis placed M3 at 44 on its Intelligence Index, well above the median for open-weight models of similar size but below the very top closed models. The fair summary: M3 is genuinely competitive on coding and agentic tasks, and it is close to the frontier rather than clearly past it.
Pricing and availability
Price is where M3 makes its loudest argument. MiniMax set standard API pricing at roughly $0.60 per million input tokens and $2.40 per million output tokens, then applied a 50% discount that brings it to about $0.30 input and $1.20 output. It went live the same day as the announcement through MiniMax’s own platform and API, and it is listed on aggregators such as OpenRouter under the identifier minimax/minimax-m3. The downloadable weights followed on Hugging Face in early June.
For comparison, that pricing sits well under the closed US frontier models and in the same low band as its Chinese open-weight peers. Cheap output tokens matter most for agentic work, where a model generates far more text than it reads.
Open-weight is not the same as open-source
This is the nuance that trips people up. "Open weight" means the trained parameters are published for download. "Open source," in the strict sense, means the license grants broad freedoms to use, modify, and redistribute, usually meeting the Open Source Initiative’s definition. Many downloadable models, M3 included, are the former but not the latter.
MiniMax released M3 under its own MiniMax Community License. Reporting at launch suggested the terms are permissive for research and general use but may require a separate agreement for large-scale commercial deployment, and some license details were still settling when the model shipped. You can read the terms on the official model card. The practical takeaway: you can download and run M3, but if you plan to build a commercial product on it, read the license first rather than assuming "open" means unrestricted. The same caution applies to self-hosting, since the weights need hundreds of gigabytes of memory and a multi-GPU setup to serve at full context.
Where MiniMax M3 sits among Chinese and US models
M3 is best understood as the latest move in a fast race among Chinese labs to ship capable open-weight models. DeepSeek set the template; its V4 Pro release showed an open model closing much of the gap with US frontier systems. Moonshot’s Kimi K3 and Zhipu’s GLM-5.2 pushed the same idea further, each competing on capability and undercutting on price. MiniMax M3 extends the pattern with its own bet: sparse attention for a genuinely usable million-token window, plus native multimodality.
Against the US frontier, the honest framing is that the closed leaders (Claude Opus 4.8, GPT-5.5, Gemini 3.1 Pro) still hold the top of most independent leaderboards, but the margin is thin and the price gap is enormous. For a cost-sensitive team doing agentic coding or long-document analysis, an open-weight model like M3 at a fraction of the price is no longer a clear downgrade. That is the real story of the 2026 open-weight wave, and MiniMax M3 is one of its sharper examples.
Frequently Asked Questions
Who makes MiniMax M3?
MiniMax is a Chinese AI lab, one of several (alongside DeepSeek, Moonshot, and Zhipu) pushing capable open-weight models onto the market. M3 is the latest release in its M-series and its most ambitious to date on context length and multimodality.
Is MiniMax M3 free to use?
The weights are free to download from Hugging Face, but running the model has real costs: either the compute to self-host it, or per-token API fees if you access it through MiniMax or an aggregator. At launch, API pricing was roughly $0.30 per million input tokens and $1.20 per million output tokens after a 50% discount. Check current rates before budgeting.
How big is the MiniMax M3 context window in words?
The context window is about 1,048,576 tokens. Because a token is a little shorter than an average English word, that works out to roughly 700,000 to 750,000 words, enough to hold a large codebase or a book-length document in a single prompt.
Is MiniMax M3 open source?
Not in the strict sense. M3 is open-weight (the trained parameters are downloadable) but it ships under the MiniMax Community License, which is not the same as an Open Source Initiative-approved license. Commercial use may require a separate agreement, so read the license before building a product on it.
Can I run MiniMax M3 on my own hardware?
Yes, if your hardware is large enough. Serving M3 at full context calls for hundreds of gigabytes of memory and a multi-GPU cluster, which puts local deployment out of reach for a single consumer machine. Most teams will access it through an API and reserve self-hosting for cases that genuinely need on-premise control.
Is MiniMax M3 better than GPT-5.5 or Claude?
MiniMax reports a narrow lead over GPT-5.5 and Gemini 3.1 Pro on the SWE-Bench Pro coding benchmark, but those results were run on MiniMax’s own infrastructure and await independent confirmation. Independent evaluators rank M3 as strongly competitive rather than clearly ahead. The closed US frontier models still lead most neutral leaderboards, though by a slim margin and at far higher prices.
What is MiniMax Sparse Attention (MSA)?
MSA is the attention mechanism at the core of M3. Instead of comparing every token against every other token (which gets expensive as context grows), it selects only the most relevant blocks of context to attend to. MiniMax says this cuts per-token compute to about one-twentieth of its previous generation at long context, which is what makes the million-token window practical to run.
