Five settings for coding work

MiniMax now lists MiniMax-M3.1-Flash-Preview in its model guide as a multimodal coding model with a one-million-token context window. The preview is available through MiniMax Code and Token Plan. Its most useful documented change is a thinking-depth control with five levels: low, medium, high, xhigh and max.

A developer could try low effort for a contained rename or a small bug fix, then raise it for a change that crosses several files and needs careful regression checks. That is a suggested way to use the control, not a measured claim about the model's speed or accuracy.

What the API documentation actually supports

MiniMax's Anthropic-compatible guide shows the exact model ID and sets thinking depth with output_config.effort. If that parameter is omitted, the preview defaults to max. Thinking cannot be switched off; requesting none or disabled returns an error. Higher settings can use more thinking tokens and add latency, so the choice has a real workflow tradeoff.

The same guide documents text, image and video inputs plus tool calls for M3.1. A team reviewing a UI bug could supply a screenshot alongside the relevant code, ask the agent to inspect the mismatch and check its proposed fix. The large context window may help with extensive repositories, but the documentation does not prove that filling it improves a particular task.

Access and what remains open

For now, MiniMax limits M3.1-Flash-Preview to MiniMax Code and Token Plan. The API guide includes an Anthropic-compatible example for eligible users, but that is not a general-access launch. MiniMax's model release-notes page has no dated M3.1 entry, and the cited pages do not provide a comparison against M3 or an independent performance result.

Start with one representative coding task and compare low and max against the same acceptance checks. Record elapsed time, token usage and whether the patch passes the tests before choosing a default for daily work.

Sources & verification

Prepared with AI assistance from the linked MiniMax documentation. XANTRANS has not independently benchmarked this preview.