MiniMax: MiniMax M1

MiniMax M1 is a text model from MiniMax, released on 17 June 2025. It is a large-scale, open-weight reasoning model built for extended context and high-efficiency inference. Under the hood it uses a hybrid Mixture-of-Experts architecture paired with a custom "lightning attention" mechanism, which is the design choice behind its long-context behaviour. If you are choosing a model for a workload that mixes reasoning with very large inputs, this is the family it belongs to.

The model is text in and text out. It accepts text prompts and returns text, with a context window of 1,000,000 tokens and a maximum output of 40,000 tokens. Its knowledge cutoff is 30 June 2024, so anything after that date has to come from the prompt rather than from the model's own training. Reasoning is part of what the model does, and the API exposes it: you can request reasoning output and inspect it through the reasoning and include_reasoning parameters rather than treating the model as a plain completion endpoint.

The obvious use cases follow from that combination. A million-token context suits work over very large documents: whole codebases, long transcripts, contract sets, research corpora or log dumps that would otherwise need chunking and retrieval. The reasoning behaviour fits tasks where the answer benefits from intermediate steps, such as multi-step analysis, structured problem solving or debugging. Tool support makes it usable in agent-style loops where the model calls functions and continues from their results. The 40,000-token output ceiling means long generations are viable, but for a single short reply it is far more headroom than most requests need.

You call it through api14's OpenAI-compatible API using the model id "minimax/minimax-m1". Because the interface is OpenAI-compatible, existing clients and SDKs work with a change of base URL and model id. The request parameters it supports are frequency_penalty, include_reasoning, max_tokens, presence_penalty, reasoning, repetition_penalty, seed, stop, temperature, tool_choice, tools, top_k and top_p. In practice that covers sampling control (temperature, top_p, top_k), penalties for repetition and presence, deterministic runs via seed, stopping conditions, output length via max_tokens, and function calling via tools and tool_choice, plus the reasoning flags.

Worth noting before you commit: input and output are text only, so image, audio or other modalities need a different model. The knowledge cutoff of 30 June 2024 means recent information must be supplied in the prompt. And a 1,000,000-token window is a capability, not an instruction; if your task is a short classification or a simple chat turn, a smaller model will usually be the more sensible fit.