MiniMax M3.1 Flash — Long Jobs On The Budget Band

October 29, 2026

MiniMax M3.1 Flash is in the model picker under Fast. Use it for a long brief, a tool-heavy client memo, or a vendor comparison when you want that work on the Budget 1× band. It holds a 1M-token context, returns up to 128K tokens, and calls tools. It does not plan a Swarm. A stronger model writes the plan, and that planner can hand MiniMax a step. Every request stays on zero-retention hosts that can run those tools.

What you can do

• Pick MiniMax M3.1 Flash under Fast. It calls tools. In a Swarm it can take a step the planner assigns. It does not write the plan.
• Keep a long job in one thread. The model reads up to 1M tokens and can return up to 128K.
• Attach a screenshot or a diagram with the brief. The model accepts text and images.
• Match the effort to the job. Fast and Medium map to low and medium reasoning. A pinned Beast asks for max.
• Stay on Budget 1×. Same band as DeepSeek V4.1 Flash and GLM 5.3 Flash.

Where this shows up

You are turning a data room into a client memo before a Monday meeting. Pin MiniMax M3.1 Flash so one thread holds the contracts, the financial model, and the recommendation, on the Budget band.

You are choosing between two vendors. Pin it and ask for a scored comparison that names the gaps each proposal left open, with the page each gap comes from.

You still have an older global MiniMax chat — M2, M2.1, M2.5, M2.7, or M3. Open it and the next turn runs on MiniMax M3.1 Flash.

Try it

• Pin MiniMax M3.1 Flash, attach the contracts and the model, then: "Write a two-page client memo: what the deal is, where the numbers do not match the narrative, and whether we should proceed."
• Pin MiniMax M3.1 Flash, attach both proposals, then: "Score these two vendors against our requirements. Show where each one is vague or silent, and tell me which questions to send before we decide."
• Pin MiniMax M3.1 Flash in Fast, then: "From this policy and last quarter's actuals, list the three gaps a buyer would challenge, each with the page it comes from."

Heads up

• Budget 1×. Same band the older global MiniMax models used, and the same band as DeepSeek V4.1 Flash.
• Zero Data Retention is required. Only hosts that keep nothing, and that can run tools, are eligible. The current zero-retention hosts for this model all support tools, so none are skipped by name.
• Check figures against the source. Artificial Analysis scored MiniMax M3.1 Flash at 42 on its intelligence index against DeepSeek V4.1 Flash's October 17 snapshot at 45, level with GLM 5.3 Flash at 42. Humanity's Last Exam was 48% versus that snapshot's 43%. Physics reasoning was 22% versus 24%. It got more factual questions right (56% versus 43%) and, when it missed, it usually gave a wrong answer instead of saying it did not know: 93% of those misses, omniscience score 16. On the newer Intelligence Index v5.0 the same lab scores it 29, with knowledge reliability at −6. Output speed there is about 118 tokens a second.
• Paid Auto does not select it. Quality 58 sits under the paid floor of 60, with GLM 5.3 Flash. Pin it when you want it every turn. Pro's Auto stays inside its curated pool either way.
• It does not plan a Swarm. The planner seat starts at quality 63 and goes to the strongest model that clears it. MiniMax can still be the worker on a step, or an Agent Smith worker outside Swarm.
• Hidden when EU Data Residency is on. This is a global route. The EU MiniMax M2.5 route on Bedrock still opens on Claude Haiku 5.5.
• Older global MiniMax chats roll forward when they are still stored on M2, M2.1, M2.5, M2.7, or M3. Chats that already moved on to DeepSeek V4.1 Flash stay there. Local MiniMax downloads are unchanged.
• No Alfrada harness numbers yet. Can it Run Alfrada? does not show MiniMax M3.1 Flash until it has a full run. The comparison above is Artificial Analysis, not that harness.