Meituan's LLM Is as Free as Its Delivery Service

"Hurry Up and Pivot to World Models"

Comrades and friends, the meta has shifted!

AI application folks are out of moves. The scales of victory have tipped back toward the foundation model companies once again. In light of this, we're paying homage to Funeral AI's series from a year ago — going through the model companies one by one.

We already covered Moonshot AI, Zhipu AI, Doubao, Qwen, and Baidu, and realized we'd shortchanged the second-tier models. So we're adding Meituan, Xiaomi, Tencent, and others in transition. More model companies coming soon 😘

Technically Meituan shouldn't be on this list, because if Meituan were a real foundation model company, its stock wouldn't be getting pummeled day after day.

But since Meitian LongCat 2.0 just dropped with all that fanfare — some people even calling it the "Zhang Xue Moment for domestic chips and domestic models" — a classic case of buzzword bingo run amok — I figured I'd do a proper postmortem on what Meituan has actually built in AI.

My first reaction to LongCat 2.0 was surprise. Didn't expect Little Wang to drop a model while simultaneously issuing his mea culpa about the stock crash being largely his fault.

What, Alibaba wasn't applying enough pressure already?

There was also that PR push about how LongCat 2.0's preview, under the pseudonym "Owl Alpha" on OpenRouter, ranked in the top three globally for API calls.

How many "top three on OpenRouter" are there? The Chinese players really need to coordinate — what if they all launched at once?

I'm saying they need to coordinate, because if everyone's simultaneously gaming OpenRouter, that's a problem.

And OpenRouter, you too — if they're going to farm you, at least charge a brand sponsorship fee next time. Don't just hand out free tokens for leaderboard spots.

This is already as routine as AI apps post-Manus: founder sitting in a WeWork speaking English, two potted plants in the background.

Is there an agency that offers package deals? AI apps get the "English + plants" package, foundation models get the "OpenRouter mutual farming" package, and both can upgrade to the "Elon Musk like" supreme package.

Another bragging point: monthly call volume ranked #1 globally on Hermes and #3 on OpenClaw, making it one of the most popular models among global Agent developers.

I asked a friend for their take. Their response: "Flexing Hermes and OpenClaw is just tasteless. Use it internally enough, throw in some free tiers, and you shoot right up."

Even the Alibaba folks don't do this anymore. The last time I remember someone running this anonymous leaderboard playbook was Happy Horse — shows how far Alibaba has come. Once their hype game reaches Moonshot AI levels, they'll come back and destroy everyone else.

Back to actual model quality.

First off, releasing LongCat 2.0 at all took guts. The model is at least better than Baidu's ERNIE Bot, since ERNIE Bot is still running gaokao benchmark scores.

Meituan's also refreshingly honest — their own benchmarks place them behind the first tier.

Funeral AI ran the full benchmark suite last week and rated LongCat 2.0 as mid-table among second-tier domestic models, sitting with DeepSeep V4 Pro, MiMo V2.5 Pro, and Seed Evolving — not as good as fellow second-tier models Hunyuan Hy3 and MiniMax M3.

Sadly, I wrote that last week. Funeral AI ran it again yesterday, and LongCat 2.0 has already dropped to the bottom of the second-tier pack, purely because competitors have been grinding too hard 😭

Worth noting: domestic models are iterating insanely fast now, basically shipping weekly. Especially Hunyuan Hy3 and MiniMax M3 — massive improvements this past month. Version numbers haven't changed, but the models have updated multiple times.

A useful reference point is Qwen 3.7 Max. Since Alibaba went all-in on Qwen 3.8 post-training and hasn't touched 3.7 for two weeks, this model that was previously on par with GLM 5.2 is now getting left behind by its peers.

So what separates the second tier?

Mainly instability. Thanks to increasingly mature distillation pipelines, these second-tier domestic models can quickly learn SOTA model action templates.

Take knowledge graph construction — second-tier models can all learn the workflow: "make a plan, read files, write single-file applications, add search and paths, run scripts, summarize."

The problem is evaluation and verification. Each of those steps requires checking actual outputs. That's where things go wrong — the model's actions look correct, but it doesn't catch problems in the actual artifacts.

So the defining trait of second-tier domestic models is instability. They can all hit high-score runs, but mostly produce low-score runs, with occasional catastrophic failures.

What can you say? Distillation isn't inherently bad. Data is what matters most in large models — human labeling is data, SOTA model outputs are data. No model lab today would claim they don't distill.

But distillation easily captures SOTA action workflows while failing to learn their evaluation and verification capabilities — that's the real dividing line between first and second tier.

How to train that capability was open-sourced in the K3 technical report published last week, and will likely appear in next week's Qwen 3.8 report as well.

Moonshot AI does solid work — they even trained the Agent runtime environment into the model. Expect second-tier domestic models to leapfrog again soon 🤓

One good thing about LongCat: internally, business units aren't forced to use it. No mandate to deploy homegrown tech — Tabbit didn't use LongCat at first either.

Another good thing: it's dirt cheap. 50 million tokens for 9.9 RMB is basically free. Shows how brutal Chinese foundation model competition is. If a US tech giant — Amazon, Microsoft, whoever — produced something like LongCat, they'd be thrilled. In China, LongCat and MiMo are practically giving it away 😭

Honestly, LongCat has gotten surprisingly far given the circumstances.

With completely inadequate resources and manpower, it's reached parity with Hunyuan, MiniMax, and Xiaomi.

This proves that foundation modeling is primarily an engineering problem. Building SOTA models requires cutting-edge technical judgment, but building second-tier models is pure engineering.

That Tencent, Meituan, and Xiaomi can all produce usable models within six months proves this is manufacturing — there's no insurmountable barrier to domestic substitution.

Which in turn validates Meituan's "domestic chips and domestic models" narrative.

LongCat is mediocre, roughly at the level of H2 2025 SOTA models.

But it trains and infers entirely on domestic chips, closing the loop on the full foundation model stack. Once the loop closes, everything else is just a matter of time — each link can be optimized incrementally, entering the steep part of the capacity curve.

The pity is Meituan's voice is nowhere near the others', fundamentally because the public perception is "they're just a food delivery company."

I have a strategy to fix this: poach Shunyu Yao to establish proper AI company credibility. Step down from that, poach his Chinese classmates. If that fails, target-acquire from their portfolio companies.

Quick inventory: Meituan is fighting on five fronts. Food delivery, international expansion, Xiaoxiang Supermarket, foundation models, and AI applications.

At least Meituan Longzhu, their investment arm, is pulling its weight — backing Zhipu AI and Moonshot AI. So the foundation model business is basically subsidized by investment returns and delivery rider sweat.

From this angle, Meituan and MiniMax are eerily similar — want everything, do everything, excel at nothing. The common read on Meituan: a company that grinds through tough businesses, with a formidable CEO who unfortunately picked the wrong赛道.

Actually Meituan's best AI product is Meituan Xiaomei. Launched to massive hype with invitation codes impossible to get — now nobody talks about it.

I use it constantly to order lemon tea and queue up for Kaojiang. Everything people wish Doubao could do, it handles — just too few people know how to use it.

Too bad Kaojiang doesn't allow online queueing — though it can call for me

Only problem: thinking way too slow. If only it could match Doubao's speed.

Later launches like NoCode, their programming tool, got absolutely destroyed by Qoder and WorkBuddy. And that Miuqu lobster social thing — pure distraction feeding CEO Xing's attention to lobsters.

As for CatPaw, their newly launched programming tool — they're attacking competitors' strengths with their own weakness.

Meituan is obviously best positioned for lifestyle Agents. Invest more in Xiaomei. Xing, stop with the random side quests.

Also, those embodied intelligence companies can barely move and still call themselves "physical AI." Meituan is the real physical AI company.

Could absolutely challenge Alibaba for who ranks #1 in physical AI. The old rivals can keep duking it out in the AI era.

I propose Meituan pivot directly to becoming a data company.

Because foundation models are pure manufacturing — same category as solar panels and EVs. Mainly about building data pipelines and continuously optimizing training capability.

That's Meituan's comfort zone. Suggest converting 100,000 riders into labeling assembly lines. Plenty of riders have solid education — filter and move them to offices. Everyone wins.

Remaining riders wear data-collection cameras on their heads. What data company could compete with that?

For real-world modeling, autonomous driving is amateur hour — Meituan is #1.

(Cover image generated by ChatGPT, text written entirely by human)

Subscribe to our Substack: funeralai.substack.com