MiniMax: MiniMax M3

Overview. MiniMax M3 is an open-weight, sparse mixture-of-experts model from MiniMax, with roughly 428B total parameters and about 23B active per token. It takes text, image, and video input and produces text output, with a 1M-token context window and up to 512K tokens of output. Developers reach for it when they need long-context, multimodal reasoning with the option to inspect the model's step-by-step thinking, without being locked into a closed-weight vendor.

Key strengths. Independent benchmarking from Artificial Analysis puts MiniMax M3's coding index at 58.6, its strongest measured area, with an intelligence index of 45.4 and an agentic index of 36.1 — a moderate, credible profile rather than a frontier one. MiniMax's own documentation describes agentic tool-use support and a configurable thinking mode (enabled, adaptive, or disabled), giving developers control over when the model exposes its reasoning trace. The model also accepts image and video input alongside text, and MiniMax reports substantial prefill and decode speed gains over its M2 predecessor thanks to a new sparse-attention design.

When to use it. Use MiniMax M3 for multi-step, tool-calling workflows and coding-assist tasks where its measured coding index is a meaningful signal, and for jobs that genuinely need its 1M-token context window — large codebases, long documents, or multimodal inputs mixing text, images, and video. Its optional thinking mode is worth turning on for tasks that benefit from a visible reasoning trace, and off for latency-sensitive, straightforward requests. It fits the 'complex' end of task routing — sustained, multi-step, or tool-using work — rather than simple single-turn queries.

Limitations. MiniMax has not published a training-data cutoff date, and the model uses a proprietary, non-standard tokenizer rather than a documented common one. Its agentic and general-intelligence index scores sit in a moderate band rather than a top tier, so it shouldn't be treated as a frontier reasoning model even with thinking mode enabled, and no independent safety or content-moderation evaluation for sensitive-domain use has been published.

Sovereignty note. MiniMax M3 is built by MiniMax, a China-based creator, and no CLOUD Act exposure, training-on-input policy, or endpoint data-residency details have been verified for this model. Because the weights are openly published under MiniMax's community licence, teams that need full control over where inference runs can self-host the model rather than relying on a third-party endpoint.