MiniMax M3.1 Flash

• 📏 1M-token context for repositories, long documents, multi-step agent runs
• 👁️ Accepts text, image and video input; returns text (catalog)
• 🧠 Always-on thinking; reasoning returned separately from the final answer
• 🔧 Five effort levels — low through max, with max as default
• 🎯 Built for agentic reasoning, tool use, coding, structured task execution
• 💬 Supports function calling and web search in the catalog listing (catalog)
• 🏢 Debuted inside MiniMax Code and the company's subscription plans
• 🆕 Newest text model in MiniMax's M-series lineup (catalog)

MiniMax

MiniMax is an AI company building generative models across multiple modalities, with a focus that spans both language understanding and audio creation. Their rapid release cadence in early 2026—delivering several new models within just a few months—reflects…

MiniMax M3.1 Flash is the newest entry in MiniMax's M-series of text models, positioned as a fast, multimodal option for coding and agentic work. MiniMax's own documentation describes it as the latest M-series language model for agentic reasoning, tool use, coding and long-context tasks, with a 1,000,000-token context window intended for long documents, whole codebases and multi-step agent sessions. The catalog listing records text, image and video input with text output, along with function calling and web search.

Within the family, it follows the M2 generation — including MiniMax M2.5 and MiniMax M2.7 — and the million-token MiniMax M3 Preview, sitting alongside MiniMax's separate media lines such as the MiniMax H3 Max video models and MiniMax Music 2.6.

The distinguishing control in M3.1 Flash is tunable thinking depth. The effort setting accepts low, medium, high, xhigh and max, with max applied when the field is omitted; higher settings produce more thorough reasoning at the cost of extra output tokens and latency. Thinking cannot be disabled — MiniMax states that reasoning before answering measurably improves accuracy on agentic reasoning, tool use, coding and maths.

Reasoning traces are returned in a separate field from the answer, so applications can display results without stripping markers out of the text. The model first shipped inside MiniMax's own coding agent and subscription plans.