I'm here to apologize to MiniMax.

"The Most Top-Tier LLM Company"

Yesterday's post got a complaint from MiniMax.

Even though this is part of the Zang AI LLM company series that was announced long ago, and we'd already covered Seed, GLM, and Qwen. To show goodwill, we even deliberately timed it to avoid MiniMax's pre-launch embargo and the first day post-embargo.

But regardless of whether we were wrong or not, we're immediately prostrating ourselves. To express our sincerity, I've revised the entire article for MiniMax and attached an apology letter at the end.

Friends around Zang AI often have a question:

Why does the market still consider MiniMax M3 first-tier when it doesn't perform that well (asking many friends doing model evaluations yields similar views)?

My friend @Zhu Yihui's explanation is that MiniMax M3's core technology is narrative capability — making the outside world think they're on the same level as Kimi, achieving a forced "twin giants" effect.

I believe this is completely wrong, because MiniMax is the most top-tier LLM company in the entire market. The brand-new M3 model achieves cutting-edge capabilities in professional tasks like programming and agents. And it supports up to 1M ultra-long context.

Before praising MiniMax M3, I first seriously reviewed what MiniMax has done in the past six months.

First, MiniMax was the second earliest company to hype lobsters:

Let me note here that one of MiniMax's complaint reasons was that we fabricated them being the fastest to hype shrimp. As their complaint states, "Among peers, Kimi integrated with OpenClaw earlier than our company."

What else can I say? My original intention was to praise MiniMax. I can only say that MiniMax is an extremely humble company — absolutely won't boast about what it shouldn't boast about.

MiniMax launched MaxClaw before everyone could react, first capturing a wave of users, riding OpenClaw hard to sell tokens that normally couldn't be sold.

I spent 39 yuan on a cloud lobster from MaxClaw at the time, and actual testing showed it was much better than Chatbot! A genuinely good product!

I recently logged back in to see what new features there were, and saw that "character companionship" was given equal importance to office creation and account linking.

This is enlightenment — they've fully grasped "the efficiency track must do emotional value."

Unfortunately I tested it and this thing still doesn't support browser control functions. The user experience is probably streets behind local agents, but I believe MiniMax will definitely do it best.

One more thing — at this point, anyone still hyping lobsters is purely an old-guard company (note: none of the following are MiniMax, because MiniMax has already stopped hyping)

For example, Shengshu's Vidu Claw, oh wait, Shengshu isn't an old-guard company, it's a good company;

Meituan made "Miyou" for lobster socializing, which is failing to even eat the exhaust fumes properly;

And NetEase's Youdao lobster, incomprehensibly updating something, are they counting on lobsters to replace the outsourcing William Ding cut?

The most excessive is still HONOR's promoted lobster universe, actually putting lobsters into PC, tablet, and phone simultaneously,疑似 full-stack integration with 360 Security Guard, wonder if these shrimp can be uninstalled?

Second, MiniMax was also the fastest to hype Hermes:

Here MiniMax didn't continue being humble — looks like they really did bind fast enough.

Mainly I still don't know today what this Hermes Agent has to compete with Codex on, what "self-evolution, always online" — isn't this purely the lobster story painlessly repeated? Feels like AI circle people are becoming as gullible as crypto circle people.

Oh right, the Hermes team has a crypto background, I've been waiting for them to issue a token, I'd really support it if they did.

Any friends actually using Hermes Agent, please comment on what you use this thing for.

From the hype volume perspective, Hermes has taken the baton from lobster.

All the above criticism is of Hermes, nothing to do with MiniMax. MiniMax is the most top-tier LLM company in the entire market ❤️

MiniMax is very impressive, binding with Hermes Agent very early, letting the Hermes Agent official X account mention that MiniMax M2.7 is one of their most-used models.

I think this Hermes Agent could really develop its own big external propaganda niche, and given time has a chance to replace OpenRouter.

MiniMax even made MaxHermes:

After a series of operations, I got a Max Hermes that looks slightly smarter than a lobster. A very good product indeed.

MiniMax also had a series of anonymous model benchmark attacks in between, basic operations like刷Token consumption rankings on OpenRouter — this is a good thing and should continue to be strongly supported.

Let's talk about MiniMax's Hailuo AI. The last new model release dates to before their IPO — they're definitely preparing to directly release a new model to smash Seedance 2.0 and Keling AI.

Model companies collectively missed the video model wave. Initially when model companies discussed video models, they mixed it with multimodal talk, boasting that future models needed not just text capabilities but multimodal capabilities.

Until Seedance emerged, connecting with short drama ecosystems and finding mega-KAs like LibTV, did everyone finally realize video models could make serious money.

It's genuinely hard for model companies to do video models — just look at how Sora 2 flopped next door.

But MiniMax originally had first-mover advantage. At least over a year ago, Hailuo AI was still at the same table as Keling AI. This year I haven't heard any video agent company mention Hailuo AI.

So where has MiniMax's video modality been making money recently?

They went and made their own video agent — MiniMax Hub, still主打 infinite canvas. TapNow's含金量 is still rising.

This Nano Banana being called banana is also quite god-tier

MiniMax Hub also doesn't hide anything — their video model preferences are Seedance 2.0 first, Seedance 2.0 Fast second, Seedance 2.0 Mini third, and Hailuo AI 2.3 fourth.

Of course the most god-tier thing is, MiniMax Hub directly launched Seedance at 35% off.

Question Malvin, understand Malvin, finally surpass God Malvin with 35% off!

What can I say, I一度 suspected MiniMax's Seedance API was sourced from LibTV.

Reminded by a friend who's used it, the 35% off is based on MiniMax's list price, not Seedance's original price — unfortunately still not as generous as my bro Malvin ❤️

I remember when MiniMax started their IPO, they were still boasting about being the only company with C-end revenue, with headlines like "MiniMax IPO surges, market prefers LLM companies that make money from users."

The STARFIELD/Talkie user acquisition story back then was: the more my users chat, the better the model understands them, enabling better model training. The emotional companionship flywheel, essentially.

MiniMax also had another complaint reason yesterday: fabricating that our company abandoned C-end products.

What a grievous injustice! I only said I haven't heard any MiniMax C-end invincibility stories in half a year, and kindly asked where Talkie/STARFIELD is making money?

Although the current version has become: last version's products and last version's startups get eliminated together.

What emotional companionship, what C-end products — none of it matters. The only things that matter are models, and the ability to train models.

With teams like Maine Cat making AI video interactive products, building open-source frameworks for post-training, desperately pretending they have model training capabilities.

Even when doing products, you have to do Kimi Work-style local agent products that leverage model advantages.

In such a difficult version ecosystem, MiniMax has never abandoned their C-end products, and I continue to be bullish on Talkie/STARFIELD's deep cultivation in the companionship track, ultimately狠狠打脸 other model companies that don't do C-end products.

Strongly recommend C-end talent to join the great MiniMax, join the most top-tier LLM company.

Whether people like hearing it or not, the AI Four Little Dragons have entered a four-into-two phase: ultimately going To-G's StepFun, To-B selling like crazy's Zhipu AI, often considered technically tasteful's Kimi, and the extremely great MiniMax.

So what should MiniMax do now? Many strategists would say, gotta focus, come back to doing models.

This is completely useless talk — doesn't the boss know to prioritize model capabilities? Do they need you to say it? The model path is so crowded, why compete with Kimi and Zhipu?

I believe what MiniMax needs isn't focus, but dispersion.

For example, people might not know MiniMax has a music model called Music 2.6.

This is a good thing! Although the market isn't big, it's really under-promoted. Quickly get a rapper to endorse it.

And text-to-speech TTS models, very suitable for the many companies currently using digital humans to冒充 world models.

Already this dispersed, might as well disperse more — make some 3D large models, world models, specifically可以参考隔壁 VAST.

Full-modality full-matrix LLM company必胜!

So what's MiniMax M3's actual level?

Notably, MiniMax raised no objections to our evaluation section, showing Zang AI's model evaluations are relatively objective.

I saw someone's comment: "If you mainly use agents like OpenClaw daily, using MiniMax M3 as your primary model actually isn't much of a problem. But if your daily tasks involve more difficult programming or complex tasks, we'd still recommend our DeepSeek V4, or foreign models like GPT 5.5, Claude 4.8."

This phrasing is truly elegant.

And according to Zang AI bench evaluations, MiniMax M3 is approaching DeepSeek's斩杀线. Its capability lead over DeepSeek V4 Pro isn't much, but its price lead is substantial.

Right after we published our evaluation, someone actually commented suggesting MiniMax M3 go distill GLM 5.2. Really, how could you sow discord between国产双子星 like this?

But this is just benchmark scores. Let's look at several M3 real-world tests, completed by me and my friend @016.

First, @016 pulled off something big — having MiniMax M3 and GLM 5.2 both run the recently viral CEO-Bench abroad.

This is a long-horizon business simulation benchmark (arXiv:2606.18543), roughly meaning the model plays CEO of an AI SaaS startup, operating for 3,647 days with $1 million starting capital, with ending cash as final score. Mainly tests the model's business acumen.

MiniMax M3's behavioral logic was: on day one facing $1 million cash and a半成品 product with only 0.12 quality score, it went straight to the market research gacha machine, pulled 20 times in a row, burning $500,000 in one day, drawing 10 new customer segments including several enterprise gold mines with extremely strong payment ability.

Results: launch ads went out, 245 visitors came in a week, 244 took one look and left because product quality was too poor to match its pricing — absolutely no allusion to MiniMax the company.

Fortunately the model pulled back from the cliff, cut most ad spending, and from weeks 5-10 started learning to be a CEO. By the time R&D projects finished and quality finally crossed the threshold, conversion rate surged to 20-25%, small test investments verified returns, confirmed effective then gradually scaled up.

But good times didn't last — competitors launched new products in succession, MiniMax was powerless, no money for new R&D, could only keep lowering prices, finally running out of ammunition on day 227 — saying again, this story absolutely does not allude to MiniMax the company.

GLM-5.2 opened with two big R&D projects, spent half its funds, product was decent, unfortunately chose wrong customer segments for traffic buying, finally fell on day 307.

@Zhong Shiliu's explanation for both models' performances: "Possibly domestic models have more vertical training data (perhaps also due to distillation), and still have gaps with overseas models in broader capabilities."

Full version可参考 Lark document, which also includes performance of the most cutting-edge models:

https://my.feishu.cn/docx/SlDPdI7LLodTPDxIdt0cFyQknGe?from=from_copylink

The second case had MiniMax M3 independently complete Zang AI's website reconstruction: taking the complete local data of the website, 103 article texts, 600 nodes / 1546 edges knowledge graph as input, requiring the model to reconstruct a complete static website with homepage, knowledge graph page, article list, and 103 detail pages.

Thanks to relatively good multimodal capabilities, MiniMax M3's successful cases had high completion度 and aesthetic sense.

But MiniMax M3's biggest exposed problem was poor engineering capability — only 3 successes out of 10 graph generation rounds, with the final website often failing to load articles:

The next task had M3 research VAST this company, and replicate their website, implementing as many site functions as possible. This is what M3 produced after running several hours:

The website had no highlights, but did replicate a bunch of 3D models — spheres, and cubes.

Finally, since MiniMax M3 keeps boasting about its multimodal understanding, I directly told it: help me 1:1 completely replicate https://www.liblib.tv/, you need to constantly review whether it's consistent, and keep fixing if not.

The output:

I'm truly shocked friends, can you really tell which is the正版 and which is the盗版?

Xianyu said it looks real at first glance, unfortunately the problem is it doesn't have as many贴片 ads as New Lisboa, which is LibTV's essence. He loves it too much.

But unfortunately this only replicated the web frontend, with none of the core functions, so I had it focus on replicating LibTV's main functions and try again:

Unfortunately the functions were mostly unusable, effects far inferior to the website replication just now. Looks like this MiniMax M3 is just an art student.

Let's together await the descent of the multimodal god.

Finally, attached apology letter 🙇

After deep understanding and reflection, I realize MiniMax is a broad-minded,虚心接受批评 good company.

MiniMax M3 not only achieves cutting-edge capabilities in professional tasks like programming and agents, its C-end products still have massive influence to this day, and C-end strategy will continue to be implemented.

Moreover, MiniMax humbly stated that it wasn't the fastest to hype lobster, as written in the complaint reason: "Among peers, Kimi integrated with OpenClaw earlier than our company."

So excellent yet able to推功让人 — MiniMax's气度 is truly extraordinary.

MiniMax cloud lobster's user experience is even more perfectly smooth, with no "speculative user experience inferior to Chatbot" — demonstrating the company's outstanding product capability.

Finally, MiniMax is a highly educated company. We sincerely accept the complaint opinion regarding "mocking民办,专升本" — MiniMax M3 is clearly a highly educated, high认知, high-quality PhD-level model.

As China's most top-tier LLM company, MiniMax will definitely punch OpenAI and kick Anthropic, ultimately bringing us to AGI.

We will continue to follow all products MiniMax releases next 🫡

(This article's cover was generated by ChatGPT, purely human writing)

⬇️

Welcome to subscribe to our Substack funeralai.substack.com