Meituan LongCat-2.0: The AI Model a Food Delivery Company Built Without Nvidia
Meituan LongCat-2.0: The Trillion-Parameter AI Model a Food Delivery Company Built Without Nvidia — and Ran Anonymously for Two Months
On June 30, Meituan — best known as China’s largest food delivery platform — launched LongCat-2.0, a 1.6-trillion-parameter AI model trained end-to-end on domestic Chinese chips. It had already been running on OpenRouter for two months under a fake name. Nobody outside the developer community noticed.
How this was reported: the facts in this article come from The Next Web’s June 30 2026 article (read in full), the VentureBeat technical analysis, and the Yahoo Tech stealth reveal article — all published June 30 2026 and read in full. Where claims are Meituan’s own assertions about their infrastructure, they are attributed as such. The specific chip manufacturer used in the training cluster has not been publicly disclosed by Meituan and is not named in this article. No interviews were conducted.
What Meituan Is and Why This Is Unexpected
Meituan is China’s largest on-demand delivery platform — food, groceries, medicine, almost anything ordered to a door in a Chinese city. It operates one of the world’s largest logistics networks. It is not, at first glance, an AI research company. Neither was Amazon when it built AWS, or Alibaba when it built Alibaba Cloud. What those companies had in common with Meituan is that running a massive logistics operation at scale requires enormous compute, and companies that depend on compute at that scale eventually decide to control it.
Meituan announced in 2025 that it would invest billions into AI and domestic chip capabilities. LongCat-2.0 is the first public result of that investment at frontier scale. The company previously released LongCat-Flash (560 billion parameters) and LongCat-Flash-Thinking as stepping stones. LongCat-2.0 at 1.6 trillion parameters is the flagship.
The Stealth Story: Two Months as Owl Alpha
The most interesting part of this launch is not the model itself — it is how Meituan released it. Before the official announcement on June 30, LongCat-2.0 had been running on OpenRouter — the AI API aggregator that lets developers access hundreds of models through a single endpoint — for approximately two months under the pseudonym “Owl Alpha.”
Independent analysis of the Owl Alpha stealth period and the LongCat-2.0 reveal, published July 2 2026. Not a Wangdoo production and not affiliated with Meituan.
During those two months, Owl Alpha accumulated real production usage from real developers who had no idea what they were using. By the time Meituan revealed the identity, the model had reached first place on the Hermes Agent workspace, second on Claude Code, and third across OpenClaw deployments, all ranked by monthly call volume. These rankings came from actual developer usage, not from curated benchmark conditions. When Meituan stepped forward on June 30, the model had already earned its position in production rather than announced it.
This is a different launch strategy from Western frontier AI releases in 2026, though it is not unique to Meituan among Chinese AI labs. PANews and other Chinese tech outlets confirm it has become a pattern: Zhipu AI’s GLM-5 ran anonymously on OpenRouter as “Pony Alpha” in February 2026, Xiaomi’s MiMo-V2-Pro ran as “Hunter Alpha” in March 2026, and Meituan’s LongCat-2.0 ran as “Owl Alpha” from late April. The naming convention — Alpha, preceded by an animal — is itself a signal that this is a coordinated approach rather than a coincidence. The strategy is straightforward: seed a model on a neutral developer platform without attaching a company name, let actual usage data accumulate, and only claim ownership once the rankings are earned.
The Chips Claim — What “End-to-End” Actually Means
The most significant technical claim in the LongCat-2.0 announcement is not the parameter count. It is the training hardware.
Meituan describes LongCat-2.0 as “the industry’s first trillion-parameter model to complete end-to-end training and inference on a 50,000-chip domestic compute cluster.” The word “end-to-end” is doing specific work in that sentence. Understanding what it means requires knowing how AI model development is divided.
Pre-training vs inference — why the distinction matters
Pre-training is the computationally intensive phase where a model processes vast datasets to learn patterns. It requires enormous parallel compute over weeks or months. This is where Nvidia’s H100 and H200 GPUs have historically dominated — they are purpose-built for this workload at scale.
Inference is the comparatively lighter task of running a trained model to answer a query. Domestic Chinese chips have increasingly been used for inference, since the compute requirements are lower and the hardware gap with Nvidia is less pronounced at this stage.
What Meituan claims — and what makes this announcement notable — is that LongCat-2.0 used domestic chips for both phases. Not Nvidia for training and domestic chips for serving. Domestic chips throughout. DeepSeek’s V4-Pro, for comparison, used Huawei chips only for inference.
That claim rests on Meituan’s own account of their infrastructure. The specific chip manufacturer or model used in the 50,000-chip cluster has not been publicly disclosed. There is no independent audit of the training hardware. The open-source community will be able to test the model’s capabilities, but the hardware provenance remains Meituan’s assertion. The Next Web noted this directly: “The training-hardware claim is harder for outsiders to confirm directly, since it rests on Meituan’s account of its own infrastructure, and that caveat is worth holding in mind alongside the company’s confidence.”
What a Mixture-of-Experts Model Actually Is
The 1.6 trillion parameter figure is real but can be misleading without context. LongCat-2.0 uses a Mixture-of-Experts (MoE) architecture, which means the model does not activate all 1.6 trillion parameters for every request. Instead, a routing system selects a subset of specialised sub-networks (“experts”) for each token processed. For LongCat-2.0, the average active parameter count per token is approximately 48 billion, ranging from 33 billion for simpler queries to 56 billion for more complex ones.
This matters for two practical reasons. First, the computational cost of running the model is much closer to a 48-billion-parameter model than a 1.6-trillion-parameter one — making it significantly cheaper to operate than the headline figure implies. Second, it means the model can specialise different expert networks for different types of tasks — coding, reasoning, long-context analysis — without the full weight of the model being active for every request. Meituan describes this as a “Zero-Compute Experts” framework that eliminates idle computational overhead on routine tasks.
The 1 million token context window is a genuine capability, not a marketing figure — it puts LongCat-2.0 on par with the longest-context models in production today and is particularly relevant for its stated positioning as an agentic coding model, where processing entire large codebases in a single pass has direct practical value.
What You Can Actually Use Today
As of July 6 2026, LongCat-2.0 is accessible through Meituan’s own API, which offers OpenAI-compatible and Anthropic-compatible endpoints. The model is also available through OpenRouter. Standard pricing is $0.75 per million input tokens and $2.95 per million output tokens, with a time-limited promotional rate of $0.30 per million input tokens and $1.20 per million output tokens.
Model weights — check current status
LongCat-2.0 is MIT-licensed. At the June 30 launch, both the GitHub and Hugging Face pages showed “Model weights coming soon — stay tuned” — confirmed from Decrypt and VentureBeat on that date. As of publication, Meituan’s Hugging Face page now hosts full model cards for both LongCat-2.0 and an INT8 quantised version, suggesting weights may now be available. Check the Hugging Face model page directly for current download availability before assuming self-hosting is possible.
For developers already using agent frameworks, LongCat-2.0 integrates directly with Hermes, Claude Code, and OpenClaw — the same frameworks where it built its two-month production record as Owl Alpha. Its architecture is specifically designed for agentic coding: long-context processing, structured output, and multi-step task execution rather than general-purpose conversation.
Meituan’s own benchmark figures give the clearest picture of where the model competes and where it does not. On SWE-bench Pro — which scores how often a model resolves real GitHub issues from production codebases — LongCat-2.0 scored 59.5, ahead of GPT-5.5’s 58.6 and Gemini 3.1 Pro’s 54.2, but behind Claude Opus 4.7 and 4.8. On FORTE — which grades AI agents on day-to-day office tasks across 15 professions under a 45-minute time limit — it scored 73.2, tied with Claude Opus 4.6 but trailing GPT-5.5’s 77.8. These are Meituan’s own reported figures, not independently audited. The pattern they show is consistent: LongCat-2.0 leads on coding-specific benchmarks and trails on broader general-agent tasks. For the pricing context that makes those results relevant: standard pricing is $0.75/$2.95 per million tokens versus GPT-5.5’s $5/$30 and Claude Sonnet 5’s $2/$10. At a fifth of GPT-5.5’s cost on a benchmark where it outperforms it, the pricing case for coding workloads is real even before the promotional rate.
Why This Connects to the Broader Chip Competition
US export controls restrict China’s access to the most advanced AI chips — primarily Nvidia’s H100 and H200 series and their successors. The policy rationale is that frontier AI capabilities require frontier silicon, and restricting silicon access restricts capability development. LongCat-2.0 is a direct challenge to that rationale.
If Meituan’s claim holds — that a 1.6-trillion-parameter model was trained end-to-end on domestic silicon — it suggests the gap between domestic Chinese chips and Nvidia hardware is narrowing faster than the export control policy assumed. The independent verification will come from the open-source community once weights are released. Until then, the claim is Meituan’s.
The context of this announcement includes China recently topping the global supercomputing rankings without US chips — LineShine debuted at number one on the TOP500 list on June 23, built entirely on domestic processors. Experts noted a specific caveat: on the HPL-MxP benchmark designed to simulate AI training workloads, LineShine ranked fourth, behind three American GPU-accelerated systems. China won the general-purpose computing benchmark; the US still leads on AI-specific compute. LongCat-2.0 sits in the same story — the claim is domestic chips can now handle frontier-scale AI training, and the independent verification is still coming. LongCat-2.0 is the AI model layer of the same domestic stack. The broader question of AI infrastructure sovereignty — which has driven government decisions from the UAE to Switzerland to California in 2026 — now has a specific Chinese answer at the model level: build it yourself.
My Take — Mr Wangdoo
The stealth launch is the detail I keep returning to. Running an anonymous model on a public API aggregator for two months and letting it earn its ranking in production is a more rigorous kind of proof than any benchmark Meituan could have published on launch day. It is also a statement: we were confident enough in this model to put it in front of real developers without telling them who built it. That is the kind of confidence that is either well-founded or about to be exposed, and two months of first-place rankings on agent workloads is at least consistent with the former.
The chips claim is harder to evaluate, and the article is honest about that. “End-to-end on domestic silicon” is the most consequential assertion in the announcement, and it is currently unverifiable by anyone outside Meituan’s data centres. When the weights arrive on Hugging Face, the community will be able to test the model. The training hardware they cannot audit. That caveat does not make the claim false — it makes it a claim.
The broader point: a food delivery company published a model that spent two months earning first-place rankings on agent and coding workloads against tools from Anthropic and OpenAI — without any of those developers knowing who built it. Whatever happens with the chips verification story, the model’s production track record is the most honest benchmark it could have had.
Frequently Asked Questions
What is Meituan and why is it building AI models?
Meituan is China’s largest on-demand delivery platform — food, groceries, and services delivered to hundreds of millions of users across Chinese cities. It runs one of the world’s largest logistics operations. Like Amazon building AWS or Alibaba building Alibaba Cloud, Meituan’s dependence on massive compute at scale made AI infrastructure a strategic necessity rather than an experiment. In 2025, the company committed to investing billions in AI and domestic chip development. LongCat-2.0 is the first public result at frontier scale.
What was Owl Alpha?
Owl Alpha was the pseudonym under which LongCat-2.0 ran on OpenRouter for approximately two months before the June 30 announcement. Meituan used this approach to accumulate real production usage data without disclosing the model’s identity or origin. By the time the reveal happened, the model had reached first place on Hermes Agent, second on Claude Code, and third on OpenClaw by monthly call volume — rankings earned from actual developer usage, not curated demos.
What does “end-to-end on domestic chips” mean and why does it matter?
AI model development has two major phases: pre-training (the computationally intensive process where the model learns from data) and inference (answering queries once trained). Most Chinese models using domestic chips have used them only for inference — the lighter task. Meituan claims LongCat-2.0 used domestic chips for both pre-training and inference. If confirmed, this means a frontier-scale model was built without Nvidia hardware, which directly challenges the premise of US export controls that restrict Chinese access to advanced AI chips.
Can I run LongCat-2.0 myself?
You can access it through Meituan’s API or OpenRouter today. The model is MIT-licensed. However, the full model weights had not been posted to GitHub or Hugging Face at the time of writing, so self-hosting is not yet possible. Meituan’s pages for the model listed “weights coming soon.” Once the weights are released, developers will be able to download, fine-tune, and self-host the model under the MIT licence terms.
How does LongCat-2.0 compare to other major models?
Meituan says performance is comparable to Google’s Gemini 3.1 Pro (released February 2026). This is the company’s own claim. The two months of OpenRouter production data show it ranked highly on agent and coding workloads against tools from Anthropic and OpenAI. Independent benchmarking by the open-source community will be possible once model weights are released. At standard pricing of $0.75 per million input tokens, it is significantly cheaper than the flagship models it claims to match on coding tasks.
Sources
- China’s Meituan says its new AI model was trained on domestic chips — The Next Web, June 30 2026 (primary source, read in full)
- Meituan open sources LongCat-2.0, the 1.6T near-frontier agentic coding model — VentureBeat, June 30 2026 (MoE architecture, pricing, weight availability status)
- LongCat-2.0: The Stealth AI Model That Was Quietly Topping OpenRouter All Along — Yahoo Tech, June 30 2026 (Owl Alpha stealth run, OpenRouter rankings confirmed)
- China claims biggest AI model trained on local chips as Meituan releases LongCat-2.0 — South China Morning Post, June 30 2026 (DeepSeek V4-Pro comparison, domestic chip context)
