01What MiniMax M3.1 Flash is and the Space Bunny story

MiniMax M3.1 Flash is a fast AI model built for agentic work, and it became known first as a mystery model called Space Bunny. Nobody knew who made it. There was no announcement and no branding, yet it sat at the top of several leaderboards.

It handled complex tasks, processed large amounts of work and beat models that had been in the spotlight for months. Developers ran 24 independent test groups to try to find out who made it.

Then came the reveal. Space Bunny Alpha is MiniMax M3.1 Flash Preview. MiniMax announced that on September 27th.

Julian Goldie says this order matters. The model did not launch with a big press push and then fade. It launched the other way around.

Developers had already used it, rated it and compared it, so the verdict was in before MiniMax put its name on it.

The rest of the episode looks at what the model can do, because the capabilities matter more than the story.

02MiniMax M3.1 Flash features: agentic work and self-checking

MiniMax M3.1 Flash is made for agentic work. That means multi-step tasks, where the AI does not just answer one question. It plans, uses tools, runs the work, checks its own output and delivers a finished result.

That is a different use from a standard chatbot.

One detail keeps coming up in independent tests: the model checks its own output. It runs something, looks at the result, and if something is off, it keeps going and corrects itself. Most models give you an answer and stop.

This one loops back.

Here is what that gives you in practice:

  • Multi-step tasks that run without constant supervision
  • Work that is checked before you see it
  • A finished result, not just a first draft

The model also supports a 1 million token context window, which Julian covers later in the episode.

03MiniMax M3.1 Flash speed and benchmark scores

Independent tests showed MiniMax M3.1 Flash hitting around 100 to 160 tokens per second, consistently. The earlier MiniMax M3 was already fast, and M3.1 Flash is measurably faster. The MiniMax team says it is nearly three times faster on agentic tasks, and independent testing so far says the speed is real.

MiniMax reports these benchmark scores:

  • Deep SWE v1.1, which tests complex software engineering such as reading code, finding issues, fixing them and running tests: 73.7%.
  • OSWorld 5.2, which measures whether a model can operate a computer environment: 61.7%.
  • CyberGym, a security benchmark about finding and validating vulnerabilities: 82%.

Julian notes these are figures reported by MiniMax. Independent testers also ran it against Claude Opus 5.5, a flagship model, and said M3.1 Flash held its own across design tasks, visual work and coding.

04Using the 1 million token context window of MiniMax M3.1 Flash

MiniMax M3.1 Flash supports a context window of 1 million tokens. The earlier M2 family topped out around 204,800 tokens. Going from that to 1 million is not just a bigger number.

It changes what you can ask the model to do.

With that much space, you can feed in large sets of material at once. Julian gives these examples:

  • An entire document archive
  • A full product database
  • Customer support emails
  • Every member onboarding form from the last year

The model can hold all of it in context at once and work across it without losing track.

This pairs well with the self-checking behavior. You can give the model a large body of material, ask for a multi-step job, and let it plan, run and check the work. The pricing, Julian says, makes high-volume automation practical, though he does not give specific prices in the episode.

05How to use MiniMax M3.1 Flash for business automation

Julian maps out a 30 day roadmap for building workflows with MiniMax M3.1 Flash inside the AI Profit Boardroom. You can use the same order on your own:

  1. Connect M3.1 Flash to your first real business task, such as qualifying leads or automating follow-up responses.
  2. Build your first multi-step agentic chain, so the model plans, executes and checks its own work without you watching it.
  3. Plug that chain into your content or onboarding pipeline so it runs on its own.
  4. Go deeper each day with content automation, lead gen, client workflows and outreach sequences.

The point is to start with one real task, not a test prompt. A model that plans and corrects itself is most useful on jobs that have several steps.

Members bring their own M3.1 Flash setups to four live coaching calls each week, and many are already running MiniMax models in their automation stacks.

06Tips and limits when testing MiniMax M3.1 Flash

Keep a few points in mind as you test the model.

  • Benchmark scores come from MiniMax. The Deep SWE, OSWorld and CyberGym numbers are reported by the company. Independent testers back up the speed and the overall quality.
  • It is a preview name. The official name MiniMax announced is MiniMax M3.1 Flash Preview.
  • Start with multi-step jobs. The model is built for agentic work, not for single questions.
  • Plan your workflow first. Julian says the gap for most people is not the model. It is the workflow, the prompts, the structure and the roadmap to deploy it.

Test it on your own data. The long context and the self-checking help most when the task is big enough to need them.

07Who MiniMax M3.1 Flash is for and why it matters

MiniMax M3.1 Flash suits business owners who want to automate lead gen, follow-up, content and onboarding. It is a fast, capable, agentic model with a 1 million token context window and self-correcting behavior on complex tasks.

Julian says the question right now is not whether tools like this work. The question is whether you are set up to use them. Many people see a powerful model but do not have the workflow, prompts or roadmap to deploy it.

Those who get that structure in place first will be running automated lead gen, follow-up, content and onboarding while others are still working out their first prompt.

Inside the AI Profit Boardroom, members get daily tutorials, a prompt library built around agentic workflows, and a member map. The AI Success Lab is the free community, with over 100 AI use cases and SOPs.

The takeaway from the episode: M3.1 Flash launched as Space Bunny, and it earned its reputation before it had a name.