MiniMax launches M3.1 Flash agent model

AI company MiniMax has officially released M3.1 Flash, a light and fast AI model built for coding and for "agents," AI helpers that carry out tasks step by step. It came out about 48 hours after DeepSeek's own Flash model, 36Kr reported. Our take: cheap, fast models like these make it less costly for anyone to build with AI.

What happened

MiniMax launched M3.1 Flash on October 19, 2026, per 36Kr, citing tech outlet Zhidx. Before the launch, the same model was being tested under the hidden name "Space Bunny."

MiniMax says the model is tuned for coding and agent work and can handle 1 million tokens at once. Tokens are small chunks of text, so that is room for very long documents or whole codebases. The company's developer guide says it accepts text, images and video, and is "designed for agentic reasoning, tool use, coding, and structured task execution."

MiniMax says it uses a "more aggressive sparse attention" design, per 36Kr. In plain terms, the model skips some work to save effort.

Key facts

• Price: 1 yuan per million input tokens and 4 yuan per million output tokens, with cached input at 0.1 yuan, per 36Kr.
• Cost claim: MiniMax says agent tasks cost 70% less overall than with its previous model, 36Kr reported.
• Speed: Over 200 tokens per second in daily use, with a typical (median) speed above 170, according to 36Kr.
• Coding test: 73.7 on DeepSWE v1.1, 0.3 points behind Opus 5. DeepSeek V4.1 Flash scored 77.0 on the same test, per 36Kr.
• Automation test: 62.2 on Automation-Bench, ahead of DeepSeek V4.1 Flash-1017 at 60.5, per 36Kr.
• Thinking control: Users can pick how hard it "thinks," from low to max, but cannot switch thinking off, says MiniMax's developer guide.

Background

DeepSeek released V4.1 Flash on October 17, 2026, about two days before MiniMax, per 36Kr. The report describes a "two strong" pattern among Chinese Flash models, with DeepSeek and MiniMax going head to head. MiniMax's input and output prices are about half of DeepSeek's peak-hour prices and match its off-peak prices.

Before launch, while it was still the hidden "Space Bunny" model, it ranked second by usage on OpenCode and once hit first place on OpenRouter, 36Kr reported.

What it means for you

Our take: you may never call this model yourself, but you will feel it. Apps that run AI helpers pay per token. When models this capable get this cheap, more small apps can afford to add AI features, and prices for users can stay low.

Our take: test scores come from the company and the report, not from us. A model that scores well on a test can still stumble on your own work, so try it on a real task before you switch.

What happens next

MiniMax is expected to release a Pro version of the model later, per 36Kr. No date was given.

FAQ

What is MiniMax M3.1 Flash?

It is a light, fast AI model from MiniMax made for coding and AI agents. It can read up to 1 million tokens at once and accepts text, images and video, per MiniMax's developer guide.

How much does MiniMax M3.1 Flash cost?

1 yuan per million input tokens and 4 yuan per million output tokens, per 36Kr.

Is it better than DeepSeek V4.1 Flash?

It depends on the task. It scored higher on Automation-Bench (62.2 vs 60.5) but lower on the DeepSWE coding test (73.7 vs 77.0), per 36Kr.

Sources

• eu.36kr.com
• platform.minimax.io