MiniMax Releases M3.1 Flash Model, ALE Score of 29.6% Ranks First Among Domestic Models

QbitAI | WeChat Official Account QbitAI

During the National Day holiday, the mysterious model Space Bunny (Jade Rabbit) that suddenly appeared on OpenRouter and anonymously topped the charts has finally had its disguise stripped off.

Beneath the rabbit skin, it turns out to be MiniMax!

The official codename upon submission is M3.1 Flash.

And this official version's benchmark performance is even a bit fiercer than Space Bunny's, especially in the Agent skill point—

It scored 29.6% on Agents' Last Exam (ALE), claiming first place among domestic models.

It also got 43 points on the AA leaderboard; for a Flash model, that's still decent.

In short, its work capability has finally come up, promoting it to a senior workhorse (doge).

Beyond that, Coding, cybersecurity, Computer Use... all have charged into the global top tier.

On top of that, the speed from the "Jade Rabbit" era has also been maintained, with output reaching up to 200+ TPS.

And then, most importantly—

As expected of Flash, the price is really cheap!!

1 yuan per million input tokens, 4 yuan for output, and cache reads cost only 0.1 yuan.

So righteous, oh—I hereby declare: in idle times, Liang Wengu; in busy times, Yan Junfeng! (bushi)

△ Image source: @Big Tech Microscope

However, Silicon Valley's closed-source camp will probably suffer again.

Following DeepSeek, a host of Silicon Valley large-parameter flagships have been put on a brand-new cost-performance kill line by domestic Flash models o(╥﹏╥)o.

To be honest, I've had quite a few doubts about MiniMax before.

Multimodal and video generation are indeed top-notch, but on the pure-text large model side, I don't know when it went quiet...

This time, with the debut of M3.1 Flash, it has staged a comeback.

Let's go through the results one by one:

Code engineering and real bug fixing

On the DeepSWE v1.1 evaluation, M3.1 Flash scored 73.7%.

That's a total reversal of the natural order.

GPT-6.1 Sol, Claude Opus 5, Claude Fable 5... actually got outscored by a Flash model.

Ahead of it, only DeepSeek V4.1 Flash 1017 remains with 77 points.

This also confirms community feedback during Jade Rabbit's anonymous testing period: when handling 3D rendering and complex front-end tasks like Three.js, its control over details is extremely obsessive.

Cybersecurity in practice

In the CyberGym benchmark led by UC Berkeley, M3.1 Flash scored 82.2 points, firmly sitting alongside GLM-5.3 (84.5) and GPT-5.6 Sol (83.6) in the global top tier.

Complex Agents and enterprise-level cross-SaaS workflows

In the AutomationBench-AA evaluation open-sourced by automation giant Zapier, M3.1 Flash ranks among the top.

With a per-task cost of less than $0.05, it can string together complex workflows spanning sales, marketing, customer service, finance, and HR on its own.

On BankerToolBench, a finance-specific benchmark with an extremely low tolerance for error, M3.1 Flash shot straight to 76.2 points, tying the recognized ceiling Claude Opus 5 for first place globally.

Computer Use real desktop operation

It scored 61.7 points on OSWorld V2, which simulates real computer operations;

On ALE, which measures real and economically valuable scenarios, M3.1 Flash also achieved top-model-level performance at an extremely low task cost.

Both rank first among domestic models—another cost-performance approach that kills everything.

Honestly, I thought MiniMax's text models would sooner or later return to the table.

But who would have thought it would be a little guy with a Flash suffix...

Looking only at benchmark scores, you might still think this is just another day of gods fighting.

The truly exciting part is this model's dimensional strike on real-world deployment scenarios.

When people usually build multi-Agent collaboration or long-chain workflows, what are they most afraid of?

Two things: endless spinning and waiting, and month-end API bills that make your heart jump.

A real enterprise-level Agent often has to call dozens of tools in the background and repeatedly pass messages among several sub-Agents.

If the model outputs slowly, users watch the interface spin and just close the webpage; if the per-call price is expensive, finance takes one look at the bill and can stop the project on the spot.

M3.1 Flash is delightful precisely here.

Under the hood, it uses a fairly aggressive sparse attention architecture.

Online inference speed is directly boosted by nearly 3x, with output speed reaching up to 200+ TPS, and the median (P50) steadily maintained at 170+ TPS.

What does 170 TPS mean?

When you dispatch a refactoring task in the MiniMax Code desktop client, Claude Code, or various local Coding tools, the code on screen almost pours out and instantly fills the screen.

Previously, chaining four sub-Agents was like waiting at traffic lights; now it's almost imperceptible refreshing.

Input and output prices are directly halved, and cache read fees plunge by 75%.

As a result, comprehensive usage cost plunges 70% compared with the previous generation M3.

So what else is there to say—stand up and pedal, folks!!

In the past, everyone had a default consensus: Flash models are nothing more than cheap castrated versions of flagship large models.

You choose Flash because you're short on money or the task is too trivial; when it comes to key, complex hardcore scenarios, in the end you still have to pay more for a large-parameter flagship Pro.

But now?

DeepSeek V4.1 Flash 1017 had just set the cost-performance benchmark, and MiniMax M3.1 Flash came right on its heels—one kills, then the other keeps killing.

Small-size models are not toys at all; they can go head-to-head with flagship models that have several times the parameters and ten times the price.

Haters, speak!!

Even more interestingly, these two domestic Flash models have taken an extremely beautiful complementary path:

DeepSeek takes the crown for extreme cost-performance in pure Coding;

MiniMax M3.1 Flash strengthens in comprehensive Agent dimensions such as cross-SaaS enterprise workflows and financial toolchain calls.

One hammers away at code engineering, the other digs deep into complex business flows.

And folks, while scrolling the internet, I found a genius idea!

Doesn't Big Fat Fish have off-peak and peak prices? But before, there was no same-tier cost-performance model, so people gritted their teeth and paid peak prices.

But now that M3.1 is out, it fills exactly that gap.

I'll just route directly: MiniMax when busy, Big Fat Fish when idle, thoroughly squeezing the Token dividend dry.

Tearing up—this is our exciting domestic Flash combo skill (bushi).

In this way, I think API is actually an option ordinary users can consider too.

After all, large models iterate really too fast, even if what you buy is just a monthly subscription.

Buying Pro 10X for nothing—my wallet hurts so much.

And C-end users are really miserable right now o(╥﹏╥)o.

There's the 5-hour limit, midnight Reset backstabs, and you constantly have to guard against behind-the-scenes intelligence-lowering sneak attacks...

Buy API, top up a few dozen yuan, and say goodbye anytime something underperforms—it's just too comfortable.

Tearing up, folks.

I don't know if everyone feels the same, but honestly, this past month I've felt a bit stifled.

Company A, OpenAI, and Google have been frantically releasing flagship models, while the domestic camp has made no moves at all.

Many big shots are also saying the LLM battle is basically over, and Silicon Valley killed everything with Scaling.

But now, an inflection point seems to be emerging.

In just two short days, two domestic Flash models have appeared one after another.

Flagship-level capability is accelerating toward universal access at Flash speed and Flash prices.

Indeed, the SOTA battle is not easy to fight.

But when the premise of "cost-performance" is satisfied, another more important battle—

The industrial deployment of large models—perhaps has only just begun.

And this has always been what Chinese companies are best at.