For about two months, the fastest-rising AI model on OpenRouter — the marketplace developers use to route billions of requests across every major model — had no name. It was listed only as "Owl Alpha." Nobody knew who built it. It quietly climbed to the number one spot for agent workloads, handling somewhere around 10 to 11 trillion tokens a month, before anyone figured out what they were actually using. Then, on June 30, 2026, the mystery ended: Owl Alpha was LongCat-2.0, an open-source model from Meituan — yes, the Chinese food-delivery and local-services giant. It was released under the permissive MIT license, meaning anyone on earth can download it, run it, and build a business on it for free. Sources: KuCoin, June 30, 2026; VentureBeat, June 30, 2026.
That reveal matters far beyond one leaderboard. LongCat-2.0 is a 1.6-trillion-parameter model that Meituan says was trained and served from start to finish on roughly 50,000 domestic Chinese AI chips — with no Nvidia hardware anywhere in the loop. If that claim holds up, it is the first frontier-scale model built entirely outside the American chip supply chain that Washington has spent three years trying to wall off. This is the honest guide for Americans: what LongCat-2.0 is, how it actually benchmarks against GPT-5.5, Gemini 3.1 Pro and Claude Opus 4.8, why the chips are the real story, what it costs, and whether you should touch a Chinese model at all.
Quick summary: LongCat-2.0 is a 1.6-trillion-parameter open-source (MIT) coding model from Meituan, with about 48 billion parameters active per token and a native 1-million-token context window. It ran anonymously as Owl Alpha on OpenRouter for roughly two months, reaching #1 for Hermes agent workloads and #2 on Claude Code deployments. On coding it is competitive with the American frontier — SWE-bench Pro 59.5, just ahead of GPT-5.5's 58.6 and Gemini 3.1 Pro's 54.2, though that gap is within eval noise and it still trails Claude Opus 4.7 and 4.8. API pricing is $0.75 per million input tokens and $2.95 output (with a launch promo of $0.30 and $1.20), roughly one-seventh the input price of a flagship like Claude Opus 4.8. Every benchmark here is vendor-reported and awaited independent verification at launch.
The Stealth Launch: How 'Owl Alpha' Fooled Everyone
Releasing a model anonymously and letting it climb the charts on merit is becoming a real strategy — and Meituan ran the playbook perfectly. For roughly two months, "Owl Alpha" sat on OpenRouter with no lab attached to it, so developers judged it purely on whether it did the job. It did. By the time Meituan claimed it, the model held the number one ranking for Hermes agent workloads, number two on Claude Code deployments, and number three across OpenClaw environments, measured by monthly call volume. Its throughput had grown to roughly 10 to 11 trillion tokens a month — a roughly 240 percent month-over-month jump — which for a coding model is an enormous amount of real, paid usage. Sources: KuCoin, June 30, 2026; geopolitechs.org, June 2026.
The stealth move was clever for a specific reason. A Chinese model carries baggage in Western developer circles — assumptions about quality, about safety, about politics. By hiding the badge, Meituan forced the market to evaluate the model blind. Thousands of developers built Owl Alpha into their coding agents and workflows because it was fast, cheap, and good, and only found out afterward that they had been running a Chinese model the whole time. That is a more convincing proof point than any benchmark chart a lab publishes about its own model.
What LongCat-2.0 Actually Is
Under the hood, LongCat-2.0 is a Mixture-of-Experts (MoE) model. The total parameter count is 1.6 trillion, but only about 48 billion are active for any given token — the model routes each request to a small slice of specialized 'experts' rather than firing all 1.6 trillion parameters every time. That is what lets a model this large run at a price this low: you pay for the compute you actually use, not the full parameter count. It ships with a native 1-million-token context window, powered by a technique Meituan calls LongCat Sparse Attention, so it can hold an entire large codebase or a stack of long documents in working memory at once.
The part that separates it from a plain open-source release is the MIT license. MIT is about as permissive as licenses get: you can download the weights from Hugging Face — full precision, plus FP8 and INT8 quantized versions for smaller hardware — run them on your own servers, fine-tune them, and ship a commercial product without paying Meituan a cent or sending it any data. That combination, frontier-scale capability plus a no-strings license, is exactly what has made Chinese open-weight models so disruptive to the American labs' business model over the past year.
China Built a Free AI on Chips Washington Banned. It Just Made $25/Month Subscriptions Look Absurd.
Chinese AI Models Are Winning in 2026: Kimi K2.5, GLM-5, Qwen 3.5 vs ChatGPT and Claude — What Every American Needs to Know
AI Reasoning Models 2026: Why the Gap Between 40% and 97% Accuracy Is Changing Everything (o3 vs GPT-5.4 vs Gemini — One Is Free for Developers)
The Benchmarks, Read Honestly
LongCat-2.0 is built to be a coding and agent model, and that is where its numbers are strongest. Here is its own reported scorecard, with the comparison figures I could verify against public sources. Read the coding lead with a grain of salt: a 0.9-point edge over GPT-5.5 on SWE-bench Pro is inside the margin of error, so the honest framing is 'roughly tied with the American frontier on coding,' not 'beats it.'
| Benchmark | LongCat-2.0 | How it compares |
|---|---|---|
| SWE-bench Pro (real coding tasks) | 59.5 | Narrowly ahead of GPT-5.5 (58.6) and Gemini 3.1 Pro (54.2); still behind Claude Opus 4.7/4.8 |
| SWE-bench Multilingual | 77.3 | Strong multi-language coding result |
| Terminal-Bench 2.1 (agentic terminal use) | 70.8 | Top-tier for autonomous command-line work |
| FORTE (general agent) | 73.2 | Trails GPT-5.5 (77.8) and Claude Opus 4.8 — weaker on broad, non-coding agent tasks |
| GPQA-diamond (PhD-level science) | 88.9 | Competitive with frontier reasoning models |
| IFEval (instruction following) | 90.0 | Reliable at doing exactly what it is told |
The pattern is clear and worth stating plainly: LongCat-2.0 is a specialist. On writing code, using a terminal, and following instructions, it is genuinely in the frontier conversation. On broad, open-ended agent tasks that are not coding — the kind of general reasoning where Claude Opus 4.8 pulls ahead — it visibly trails. If your work is software, it is a serious option. If you need an all-purpose brain, the American flagships still lead. And every one of these numbers came from Meituan; independent labs had not finished re-running them at launch, so treat them as a starting point, not gospel. Sources: Hugging Face model card; MarkTechPost, July 5, 2026; creativeainews.com, 2026.
The Real Story Is the Chips
The benchmarks are impressive, but they are not the headline. The headline is that Meituan says it trained this entire 1.6-trillion-parameter model — 35 trillion-plus tokens of training — on a cluster of roughly 50,000 domestic Chinese AI accelerators, with zero Nvidia GPUs, and completed the run with no rollbacks or unrecoverable loss spikes. Training a model this large is notoriously unstable; doing it cleanly on a non-Nvidia stack that most of the industry considered second-tier is a serious engineering result. Sources: South China Morning Post, June 30, 2026; MarkTechPost, July 5, 2026.
One honest caveat: Meituan did not officially name the chip. Most analysts believe it is Huawei's Ascend line — the Ascend 910C is the usual guess — but that is inference, not a confirmed fact, and you should treat any specific chip name you see reported as unverified. What is not in doubt is the strategic point. The entire premise of US export controls since 2022 has been that cutting China off from Nvidia's best chips would keep it a generation or two behind on frontier AI. LongCat-2.0 is a direct, public rebuttal: a top-of-the-charts model built without a single restricted American chip, then given away for free under an MIT license. Whatever you think of the politics, the technical claim lands.
What It Costs and How to Actually Use It
There are three ways to use LongCat-2.0, depending on how much control you want. The simplest is through OpenRouter or Meituan's own API, which is compatible with the OpenAI and Anthropic request formats, so most existing code works with a one-line endpoint change. Standard pricing is $0.75 per million input tokens and $2.95 per million output tokens, with cached context reads free; at launch there was a promotional rate of $0.30 and $1.20. To put that in perspective, a flagship proprietary model like Claude Opus 4.8 costs about seven times more per input token (roughly $5 versus $0.75 per million) — which is the entire reason a coding startup running millions of calls a day would even consider switching.
The second way is to self-host. Because the weights are MIT-licensed and published on Hugging Face, including quantized FP8 and INT8 versions, a team with its own GPUs can run the model on private infrastructure so that no prompt ever leaves the building. The third way, for individuals, is simply to use it inside a tool that already offers it — many multi-model platforms add new open-weight models within days of release. The point is that access is not the barrier here; the barrier, if there is one, is trust.
Should Americans Use a Chinese Model? The Honest Take
- Self-hosting the open weights is the safe path. Because LongCat-2.0 is MIT-licensed and downloadable, you can run it entirely on your own machines. In that setup, no data goes to China or anywhere else — it is functionally the same privacy posture as any local model. For sensitive code or business data, this is the only version worth considering.
- The hosted API sends your prompts to servers in China. Using Meituan's API (or a router that forwards to it) means your text is processed under Chinese jurisdiction and data-governance rules, which offer no equivalent to US or EU legal protections. Fine for hobby projects and non-sensitive tasks; not appropriate for proprietary source code, customer data, or anything regulated.
- The weights themselves are not a spying tool. A downloaded, self-hosted model cannot phone home — it is just math running on your hardware. The realistic open question researchers raise is subtler: whether any model's outputs could be subtly biased, which is impossible to fully rule out for any model, American or Chinese, without training-data access. For normal coding work this is a theoretical concern, not a practical one.
- Never use any Chinese-hosted model for government, defense, or national-security work. That line is not about probability; it is about standard. If your work touches classified or export-controlled material, the entire category is off-limits regardless of how good the model is.
What This Means for the AI Race
Zoom out and LongCat-2.0 is a data point in a bigger trend that should worry the American labs more than any single benchmark. Chinese labs have decided that the way to win is not to build a better paid product than OpenAI or Anthropic, but to give away models that are 90 to 100 percent as good for free, and let the world's developers build on top of them. Every time one of these models tops a neutral leaderboard while anonymous, it proves the strategy works. The American advantage in raw capability is real but shrinking, and it is being attacked from a direction — price and openness — where a $10-billion research lab has no natural defense.
For you as a user, the takeaway is simpler and more optimistic: the competition is ferocious, and it is driving the price of very good AI toward zero. The smart move is not to marry one lab or one country's models, but to stay able to pick the best tool for each job — the frontier American model when the stakes are high, a cheap open-weight model when volume matters, and whatever tops the leaderboard next month when it inevitably changes. A multi-model platform like LumiChats exists for exactly this: access to 40-plus models, American and open-source, under one login instead of a stack of $20-a-month subscriptions, so you are never locked into yesterday's winner.
01What is LongCat-2.0?
It is a 1.6-trillion-parameter open-source coding model from Meituan, the Chinese local-services company, released under the MIT license on June 30, 2026. It uses a Mixture-of-Experts design with about 48 billion parameters active per token and a 1-million-token context window.
02Was it really the secret 'Owl Alpha' model?
Yes. For roughly two months it ran anonymously as 'Owl Alpha' on OpenRouter, climbing to #1 for Hermes agent workloads and #2 on Claude Code deployments and reaching about 10 to 11 trillion tokens a month before Meituan revealed it was the model.
03Is it better than GPT-5.5, Gemini 3.1 Pro, or Claude Opus 4.8?
On coding it is competitive: SWE-bench Pro 59.5 versus GPT-5.5's 58.6 and Gemini 3.1 Pro's 54.2, though that lead is within eval noise and it still trails Claude Opus 4.7 and 4.8. On broad, non-coding agent tasks it clearly trails Claude Opus 4.8. It is a coding specialist, not an all-purpose leader.
04What is the big deal about the chips?
Meituan says it trained the entire model on roughly 50,000 domestic Chinese AI chips with no Nvidia hardware — a direct challenge to US export controls designed to keep China behind on frontier AI. The specific chip was not officially named, though many analysts assume Huawei Ascend.
05How much does it cost and how do I use it?
Standard API pricing is $0.75 per million input tokens and $2.95 output (launch promo: $0.30 and $1.20), roughly one-seventh the input price of Claude Opus 4.8. You can use it via OpenRouter, Meituan's OpenAI-compatible API, or by self-hosting the MIT-licensed weights from Hugging Face.
06Is it safe to use a Chinese AI model?
If you self-host the open weights, no data leaves your machine, so the privacy posture is the same as any local model. Using the hosted API sends your prompts to servers in China under Chinese law, which is fine for non-sensitive work but not for proprietary, regulated, or government data.
The bottom line: LongCat-2.0 is not proof that China has 'won' AI — on general capability, the American frontier still leads. It is proof of something more durable. A Chinese company built a genuinely competitive frontier model, on banned-market chips, and gave it away for free, and thousands of American developers were happily using it before they even knew where it came from. That is the shape of the race now, and pretending otherwise helps no one. Use the open weights if the price is worth it to you, self-host anything sensitive, and keep your options open — because next month there will be another Owl Alpha.
