Models
What Is MiniMax M3.1 Flash? The Space Bunny Model, Explained
MiniMax M3.1 Flash launched October 19, 2026. What it is, the $0.15 price, its agent benchmarks, the Space Bunny preview, and how it compares to DeepSeek.
MiniMax M3.1 Flash is a new AI model from the Chinese company MiniMax, released on Monday, October 19, 2026. If you saw "MiniMax M3.1" or "M3.1 Flash" in the news today, or heard developers talk about a mystery model called Space Bunny, they are talking about the same thing.
This page explains what MiniMax M3.1 Flash is, what it costs, how well it does on tests, how you can try it, and how it stacks up against DeepSeek’s cheap model that came out two days earlier.
What is MiniMax M3.1 Flash?
MiniMax is an AI company from China. Its shares trade in Hong Kong, and it makes text, voice, music and video models. Its main text model line is called M3.
M3.1 Flash is the newest model in that line. "Flash" is the label AI labs use for the quick, cheap version of a model. Google and DeepSeek use the same word. So think of M3.1 Flash as the speedy workhorse, not the biggest brain.
MiniMax built it for agents. In its launch post, MiniMax points to gains in agent tasks, programming and cybersecurity. Reuters reported the same three areas, quoting MiniMax’s statement that the gains are "generational." Agents are AI that plan steps and do the work, instead of only answering one question. If that idea is new, start with what agentic AI is.
MiniMax also says a bigger, smarter version called M3.1 Pro is coming later. It has not given a date.
What was Space Bunny?
Before the launch, MiniMax tested the model without its name on it. Developers found a free model called Space Bunny on the OpenRouter platform in late September and spent weeks guessing who made it. Many guessed MiniMax.
They were right. MiniMax now confirms that Space Bunny was an early version of M3.1 Flash. It says Space Bunny handled more than 200 trillion tokens while it ran anonymously. A token is a small piece of text, so that is a huge amount of real use before launch day. A preview version, M3.1-Flash-Preview, also showed up inside MiniMax’s coding tool on September 27.
MiniMax M3.1 Flash price
Developers pay per million tokens. MiniMax lists one flat price, with no peak hours.
| Per 1M tokens (USD) | MiniMax M3.1 Flash |
|---|---|
| Input | $0.15 |
| Cached input | $0.015 |
| Output | $0.60 |
That is very cheap. MiniMax says total cost on agent tasks is about 70% lower than its previous model. It also says normal output runs at more than 200 tokens per second, with a typical (median) speed above 170.
MiniMax has not published the model’s size or its context window in the launch post. Some trackers list numbers, but until MiniMax confirms them, treat them as unconfirmed.
MiniMax M3.1 Flash benchmarks
Here are the scores from MiniMax’s launch post. These are MiniMax’s own tests. Independent results usually take a few days to a few weeks.
| Test | What it checks | M3.1 Flash score |
|---|---|---|
| CyberGym | Security tasks | 82% |
| BankerToolBench | Using finance tools | 76.2 |
| DeepSWE v1.1 | Fixing real code | 73.7% |
| OSWorld V2 | Using a computer like a person | 61.7 |
| Agents’ Last Exam | Long multi-step agent tasks | 29.6 |
MiniMax says M3.1 Flash ranks first among Chinese models on OSWorld V2 and Agents’ Last Exam. It does not name the exact rivals in the text, so read "best among Chinese models" as MiniMax’s claim, not a settled fact.
MiniMax M3.1 Flash vs DeepSeek V4.1 Flash
The closest rival is DeepSeek V4.1 Flash, which got an update on October 17. Both are cheap Chinese "Flash" models aimed at agents. Here is how their public numbers line up.
| MiniMax M3.1 Flash | DeepSeek V4.1 Flash (1017) | |
|---|---|---|
| Released | Oct 19, 2026 | Oct 17, 2026 (first version Sep 10) |
| Input price | $0.15, all day | $0.15 off-peak, $0.30 peak |
| Output price | $0.60, all day | $0.60 off-peak, $1.20 peak |
| DeepSWE v1.1 | 73.7% | 77.0 |
| Agents’ Last Exam | 29.6 | 35.4 |
| Model files | Promised "later" | Free now, MIT license |
The short read: MiniMax charges DeepSeek’s off-peak price around the clock, which is simpler for businesses. DeepSeek posts higher scores on the two tests both companies report, and its files are already free to download. Both sets of scores come from the companies themselves, so the real gap may be different.
How to try MiniMax M3.1 Flash
MiniMax says M3.1 Flash is available in three places today:
- MiniMax API: for developers building their own apps, at the prices above.
- M Plan: MiniMax’s subscription plan.
- MiniMax Code: its coding agent tool, which runs as a desktop app.
Is MiniMax M3.1 Flash free? Not in general. You pay through the API or the plan. MiniMax says it will release the model weights later, which would let people run it on their own machines, but it has given no date and no license yet. Until that happens, "open source" is a promise, not a fact.
What this means for you
This is the third big model news story in a week, after Google’s Gemini 4 Argon and DeepSeek’s update. The pattern is clear. AI labs now sell models as workers that finish tasks, and they compete hard on price. That is good news, because cheaper and faster models end up inside the apps you use.
But a model on its own is an engine with no car. To get value from M3.1 Flash today, you need an API key, a coding tool, or a subscription, plus time to set it all up. Most people just want the task done.
That is what our app is for. Agentic AI puts a team of expert agents on your iPhone and iPad, like a Research Analyst, a Writing Coach, and a Life Planner. You describe the job, pick one agent or up to 4 working together, and get a finished result. There is no account to create, and your first 3 tasks are free. You skip the model news cycle and keep the results.
Quick answers
When was MiniMax M3.1 Flash released? MiniMax launched it on October 19, 2026. A preview version appeared in MiniMax Code on September 27, and an anonymous test version called Space Bunny ran on OpenRouter before that.
Is Space Bunny MiniMax M3.1 Flash? Yes. MiniMax confirmed that Space Bunny was an early version of M3.1 Flash, and says it processed more than 200 trillion tokens while anonymous.
How much does MiniMax M3.1 Flash cost? $0.15 per million input tokens, $0.015 per million cached input tokens, and $0.60 per million output tokens, per MiniMax’s launch post.
Is MiniMax M3.1 Flash open source? Not yet. MiniMax says the model weights will be released later but has not given a date or license.
What are the MiniMax M3.1 Flash benchmark scores? MiniMax reports 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, 29.6 on Agents’ Last Exam, 82% on CyberGym, and 76.2 on BankerToolBench. These are self-reported.
Is MiniMax M3.1 Flash better than DeepSeek V4.1 Flash? On the two tests both companies report, DeepSeek’s October 17 version scores higher. MiniMax’s pricing is simpler, with the same low rate all day. See the table above, or our agentic AI glossary for the terms.
Sources: MiniMax: M3.1 Flash launch post · Reuters via The Standard · Reuters via US News · Startup Fortune · DeepSeek API pricing
