Most AI models have a context window — a limit on how much text they can process at one time. Claude handles up to 200,000 tokens. GPT-4o caps at 128,000. MiniMax M3, released in June 2026, has a context window of 1 million tokens.

One million tokens is roughly 750,000 words. That is every email you have sent in the past two years, every client file in a project, or the full transcript of a year of meetings, fed to an AI model in a single pass.

Why Context Size Is Not Just a Technical Spec

Context limits are a practical constraint that most people only notice when they hit them. You want to ask the AI a question about a 90-page contract but the model can only see 60 pages at a time. You want to analyse a year of client communication but you have to chop it into chunks and lose the connections between them. You want to compare three competing proposals but the combined length exceeds the window.

A 1 million token window makes those constraints disappear for almost any document-based task a business might have. Load the entire project into the model and ask it anything, without managing chunks or losing context between calls.

What MiniMax M3 Is

MiniMax is a Chinese AI lab. M3 is open-weight, meaning the model weights are publicly available to download and run. It is deployable via Ollama locally or via the MiniMax API. On coding and reasoning benchmarks it sits in the same general bracket as the mid-tier paid models.

The 1 million context at open-weight terms is the differentiating factor. Paid models that offer extended context charge for it — a long-context GPT-4o request costs more per call than a standard one. A 1 million token query via a paid API at current rates would be expensive. Running M3 locally removes that variable entirely.

One Thing to Consider

MiniMax is a Chinese company. For South African businesses processing sensitive client data, the question of how the API routes data and what the lab's data practices are is worth asking before using the cloud version. Running the model locally via Ollama sidesteps this entirely — no data leaves the device. For document-heavy work on your own internal information, the local route is both the private and the cost-effective one.