MiniMax Releases M3.1 Flash, Scoring 29.6% on Agents' Last Exam to Rank First Among Chinese Models

36Kr Auto Digital Kr Future Consumption Intelligent Emergence Future City Launch Power on 36Kr Global 36Kr Research Institute Tide 36Kr Enterprise Service Review 36Kr Finance Workplace bonus 36 Carbon Post-Wave Research Institute Dark Waves Hard Kr Kr Rui Research Institute Media Brand

Enterprise Accounts Enterprise Service Review 36Kr Research Institute 36Kr Innovation Consulting Enterprise Services

Core Services City Window Government Services

Venture Capital Release LP Source Plan VClub VClub Investment Institution Database Investment Institution Job Referral Investor Certification Investor Services

Event Search Project Recommendation I Want to Join City Cooperation Domestic Model Agent No. 1, MiniMax Radically Overhauls Flash Combat Power, All-Day Big Fat Fish Valley Price Quantum Bit · October 20, 2026 20:33 Under the rabbit skin, it turns out to be M3.1 Flash! MiniMax Equity Financing Shanghai 2021-11 AI Technology R&D Provider 36Kr Report IoT/Hardware Smart Hardware WISE2025 King of Business Annual AI Application Scenario Breakthrough Enterprise WISE2024 Annual Enterprise List - Artificial Intelligence I Want to Contact

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.

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

The official submission code name is M3.1 Flash.

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

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

On the AA leaderboard it also got 43 points; for a Flash to achieve this is still quite good.

In short, its working capability has finally come up, promoted to senior workhorse (doge).

In addition, Coding, cybersecurity, Computer Use... have all also charged into the global first tier.

On this basis, the speed from the "Jade Rabbit" period has also been maintained, with output reaching up to 200+ TPS.

Then, the most important part is here—

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

Input 1 yuan per million tokens, output 4 yuan, cache read only 0.1 yuan.

So righteous, oh—I announce that from now on, in idle times Liang Wengu, in busy times Yan Junfeng! (bushi)

△

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-effectiveness kill line by domestic Flash o(╥﹏╥)o.

Text Model Comeback Battle

To be honest, I used to have quite a few misgivings about MiniMax.

Multimodality and video generation are indeed top-notch, but on the pure text large model side, I don't know when it started, there was no movement...

This time M3.1 Flash's appearance can be considered a comeback battle.

Let's look at the battle results one by one:

Code Engineering and Real Bug Fixing

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

It really turned the tables.

GPT-6.1 Sol, Claude Opus 5, Claude Fable 5... were unexpectedly surpassed in benchmark scores by a Flash.

Ahead of it, only DeepSeek V4.1 Flash 1017's 77 points remain.

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 abnormal.

Cybersecurity Practical Combat

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

Complex Agents and Enterprise-Level Cross-SaaS Workflows

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

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

On BankerToolBench, a finance-specific benchmark with extremely low fault tolerance, M3.1 Flash directly soared to 76.2 points, tied for global first with the recognized ceiling Claude Opus 5.

Computer Use Real Desktop Operation

OSWorld V2, which simulates real computer operation, got 61.7 points;

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-effectiveness method that kills everything.

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

What is truly exciting is this model's dimensional strike on actual 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 one's heart jump.

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

The model generates text slowly, and users watch the interface spin before closing the webpage outright; the per-call price is expensive, and finance can halt the project on the spot after one look at the bill.

That is exactly where M3.1 Flash feels great.

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

Online inference speed is directly boosted by nearly 3x, output speed can reach up to 200+ TPS at its fastest, and the median (P50) also steadily stays at 170+ TPS.

What does 170 TPS mean?

When you dispatch a refactoring task in the MiniMax Code desktop app, Claude Code, or various local coding tools, the code on the screen almost instantly fills up as if pouring out.

Previously, chaining four sub-Agents felt like waiting at traffic lights; now it is almost imperceptible refresh.

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

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

People chose Flash because they were short on money or the task was too trivial; when it came to key, complex, hardcore scenarios, they ultimately still had to pay more for the large-parameter flagship Pro.

But what about now?

DeepSeek V4.1 Flash 1017 had just set the cost-effectiveness benchmark, and MiniMax M3.1 Flash followed right on its heels, one cutting down after another.

Small-sized models are not toys at all; they can absolutely 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 elegant complementary path:

DeepSeek takes the lead in extreme cost-effectiveness for 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 family, while scrolling the internet, I found a genius idea!

Doesn't Big Fat Fish have valley pricing and peak pricing? But previously it suffered from never having a model of the same tier and cost-effectiveness, so even peak pricing had to be gritted through.

But now that M3.1 is out, it fills the spot perfectly.

I directly set up one routing: MiniMax when busy, Big Fat Fish when idle, thoroughly eating up every bit of the Token dividend.

Tearing up, this is our exciting domestic Flash combo skill (not).

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

After all, this large-model stuff iterates really too fast, even if what you buy is only 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 is the 5-hour limit, the midnight Reset backstab, and having to guard against backend intelligence-lowering sneak attacks at every turn...

Buy API, top up a few dozen yuan, and say goodbye anytime whichever one underperforms—it is simply too comfortable.

Tearing up, family.

I do not know if everyone has the same feeling, but to be honest, this past month I have felt a bit stifled.

Company A, OpenAI, and Google are madly releasing flagship models, while on the domestic camp side, there is no movement at all.

Many big shots are also saying that 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 price.

Indeed, the SOTA battle is not easy to fight.

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

large-model industry deployment—perhaps has only just begun.

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

This article comes from the WeChat public account "QbitAI" (ID: QbitAI), author: Focus on Frontier Technology, published by 36Kr with authorization.

The views in this article represent only the author himself, and the 36Kr platform only provides information storage space services.+123

A good article needs your encouragementQbitAI Contributing Author5Favorite+10CommentOpen WeChat "Scan", open the webpage and click the share button in the upper right cornerWeiboImmersive readingBack to topReportJoin the commentsTens of millions of comments, friendliness firstLog in to join the discussionSubmit comment0/1000You may also like these articlesApprenticeship domestically, high prices overseas: Physical AI going global, why is it more dangerous the more efficiency is discussed?September: 2026 China AI Large Model Platform RankingAltman proposes three "endgame paths" for AI development: from international regulation to "one company dominating"Behind AI computing centers, a "refrigeration war" worth hundreds of billionsThis year's AI has driven telecom scammers crazyDeducting non-recurring losses of 760 million, doubling year-on-year, iFlytek still has to wait to make money from AIBreaking: DeepSeek heavily open-sources its core inference engineAfter four rounds of hundred-million-level financing in half a year, where does Ola Dimensional want to take home robots?Computing power rents fluctuate violently, and AI also needs to buy "insurance"QbitAI Contributing Author

TA has not written a bio, but being reserved is also a form of expressionPublished 4101 articles

Recent contentDomestic model Agent No. 1, MiniMax drastically overhauls Flash combat power, all-day Big Fat Fish valley price1 hour agoThree key pieces of the RSI puzzle, MemT 1.0 fills in "memory" from the model layer6 hours agoMusk spends 10.5 billion dollars, 72 million children across the US, one share of SpaceX each10 hours agoRead more content, poke here hardView AI reviewsMiniMaxDeepSeekChoose reliable AI, look at real reviewsView 36Kr AI Review Official Communication CommunityJoinConsult project review and onboardingContact 36Kr project recommendationsSubscription accountFollow MiniMaxI want to contactArtificial intelligence technology developer

36Kr Europe36Kr Europe36Kr EuropeAI Product DailyOnline rumor information reporting entranceHot recommendations

Push and interpret cutting-edge, substantive technology venture capital information

Primary market financial information and system service provider

Focus on outstanding global entrepreneurs, project financing rate close to 97%, leading the industry