Breaking

MiniMax Brings M3.1-Flash-Preview to MiniMax Code With Tunable Reasoning Effort

September 28, 2026 at 10:19 EDT

Software Dev & Coding
AI Agents
Foundation Models

MiniMax has made MiniMax-M3.1-Flash-Preview available for coding work in MiniMax Code, positioning the preview model for everyday development tasks from bug fixes to feature implementation. The launch moves the company's M-series coding stack toward a speed-and-effort model: developers can choose how deeply the model reasons, rather than treating a coding agent as a single fixed setting.

September 27, 2026 · MiniMax

M3.1-Flash-Preview shifts coding models toward speed tiers

MiniMax has put its new preview model into MiniMax Code with a 1,000,000-token context window and five selectable reasoning-effort levels for coding-agent work.

Context window

Effort levels

400

Error if thinking is disabled

Context window jump

MiniMax's current M-series table puts M3.1-Flash-Preview at the same 1,000,000-token scale as M3 and far above the 204,800-token M2.7 line.

204,800M2.7

1,000,000M3.1-Flash-Preview

How the preview is meant to be used

Pick a task

Bug fix, feature work, tests or change verification.

→

Set effort

Choose low, medium, high, xhigh or max.

→

Run the agent

Trade latency and token use for deeper reasoning.

Positive read

Developers are testing it as a faster everyday coding model for multi-step tasks and real repository work.

Open questions

Public model-specific benchmarks, a full model card and broader release details are still missing.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.