Meituan has open-sourced LongCat-2.0, a 1.6 trillion-parameter Mixture-of-Experts model with a native 1-million-token context window, released on June 30, 2026 under a permissive MIT license. It is the model that quietly topped global developer charts on OpenRouter for two months under the codename “Owl Alpha”, and Meituan says it is the first trillion-parameter model trained and served end-to-end on Chinese-made chips, with no Nvidia GPUs involved.
This article breaks down what LongCat-2.0 actually is, how its unusual architecture keeps a 1.6T model cheap to run, what the vendor-reported benchmarks claim, and how you can try it today. We also flag where the hype outruns the evidence, because the headline numbers come from Meituan’s own testing and the full model weights are not public yet.
The Key Takeaways
- LongCat-2.0 is a 1.6T-parameter open-source MoE model that activates only about 48B parameters per token (dynamic range 33B–56B).
- It ships with a native 1M-token context window and is purpose-built for agentic coding and tool use.
- On the vendor’s own benchmarks it scores 59.5 on SWE-bench Pro, edging out GPT-5.5 (58.6), plus 70.8 on Terminal-Bench 2.1.
- It was trained on a 50,000-chip domestic ASIC cluster, the first trillion-parameter model built without Nvidia hardware.
- Promotional API pricing starts at $0.30 per million input tokens with free context-cache hits, a fraction of GPT-5.5 costs.
What Is LongCat-2.0?
LongCat-2.0 is a large language model from Meituan, the Chinese delivery and local-services giant that has been quietly building a serious AI lab. It uses a Mixture-of-Experts design with 1.6 trillion total parameters, but it only activates a small slice of them on any given token, averaging around 48 billion active parameters. That sparsity is the whole trick, because it lets the model carry frontier-scale knowledge while running at the cost of a much smaller dense model.
The model targets agentic coding specifically, meaning code understanding, generation, and multi-step execution inside agent workflows rather than one-shot chat answers. It launched as fully open source under the MIT license, which allows commercial use, and it joins a growing wave of Chinese open-weight releases sitting alongside models like GLM and Seed 2.1 Pro. If you want the wider context on the trend, our guide to open source AI covers what “open” really means here.
One caveat up front. As of launch, the Hugging Face and GitHub pages are live but list the weights as “coming soon,” so the practical way to use LongCat-2.0 right now is through hosted APIs rather than a local download.
The “Owl Alpha” Backstory
Before it had a name, LongCat-2.0 ran anonymously on OpenRouter as “Owl Alpha” for roughly two months. During that stealth run it became one of the most-used models on the platform, accounting for about 10.1 trillion monthly tokens, an average of 559 billion tokens a day, and a reported 242% month-over-month jump in volume.
That matters because it is a rare case of real-world usage data preceding the marketing. Developers were reaching for the model heavily before anyone knew Meituan was behind it, which is a stronger signal than a launch-day benchmark chart. Meituan says the architecture had already climbed to the top of several agent leaderboards and sat near the top on Claude Code deployments during that period.
How LongCat-2.0 Is Built
The reason a 1.6T model can be cheap enough to run at OpenRouter scale comes down to three architectural choices. Each one attacks a different bottleneck, namely attention cost, wasted compute, and task specialization.
LongCat Sparse Attention (LSA)
Standard attention compares every token to every other token, so cost grows quadratically as context gets longer. LongCat Sparse Attention instead selects only the most relevant tokens to attend to, which drops the scaling closer to linear. That is what makes the full 1-million-token context practical rather than ruinously slow, letting the model pull a specific function out of a giant codebase without re-reading everything.
Zero-Compute Experts
In most MoE models every token still pays a baseline cost. Zero-Compute Experts let simple tokens route through near-empty subnetworks that consume almost no compute, while genuinely hard tokens get more experts assigned to them. This is why the active parameter count flexes between 33 billion and 56 billion per token instead of staying fixed, so easy boilerplate is cheap and tricky logic gets the horsepower.
MOPD and Three Expert Groups
LongCat-2.0 splits its training into three specialized expert families fused through a method Meituan calls MOPD (Multi-Teacher On-Policy Distill). The Agent Experts handle tool use and self-correction, the Reasoning Experts handle multi-hop reasoning and adaptive computation, and the Interaction Experts handle instruction following and hallucination suppression. At inference a gating network routes each task to the experts best suited to it, rather than blending everything into one averaged model.
Benchmarks: How LongCat-2.0 Performs
Here is the part to read carefully. Every number below is vendor-reported by Meituan, not independently verified, and the much-repeated “beats GPT-5.5” line is true on exactly one benchmark by under a point. Treat these as a claimed ceiling rather than a settled ranking.
| Benchmark | What it tests | LongCat-2.0 | Note |
|---|---|---|---|
| SWE-bench Pro | Resolving hard real-world GitHub issues | 59.5 | Edges GPT-5.5 (58.6) |
| Terminal-Bench 2.1 | Agentic command-line task completion | 70.8 | Coding agent strength |
| SWE-bench Multilingual | Software fixes across many languages | 77.3 | Broad language coverage |
| FORTE | Enterprise workflow simulation | 73.2 | General agent tasks |
| RWSearch | Real-world agentic search | 78.8 | Tool-augmented retrieval |
| BrowseComp | Hard web-browsing research questions | 79.9 | Deep research tasks |
The honest read is that LongCat-2.0 lands in near-frontier territory for agentic coding, which is genuinely impressive for an open-weight model, while the GPT-5.5 comparison is a marketing-friendly framing of a statistical tie on one test. The browsing and search scores are the standout numbers and line up with the heavy agent usage it saw as Owl Alpha.
Trained Without Nvidia
The detail drawing the most attention is the hardware. Meituan says LongCat-2.0 was trained and run on a cluster of more than 50,000 Chinese-made ASICs, with no Nvidia GPUs in the loop, which it claims makes this the first trillion-parameter model fully built on domestic compute.
That is a meaningful data point in the ongoing chip-export story. It suggests Chinese labs are closing the gap on training-scale infrastructure even under hardware restrictions, and it removes a supply-chain dependency that has constrained other developers. For the broader picture of where these open-weight contenders stack up, see our roundup of the best open source AI models.
Pricing and How to Access It
Cost is where LongCat-2.0 gets aggressive. Standard API pricing runs about $0.75 per million input tokens and $2.95 per million output tokens, and a launch promotion cuts that to roughly $0.30 input and $1.20 output, with context-cache hits billed free. For comparison, GPT-5.5 sits near $5 input and $30 output per million tokens, so the gap for heavy agent workloads is large.
| Model | Input (per 1M) | Output (per 1M) | Cache hits |
|---|---|---|---|
| LongCat-2.0 (promo) | $0.30 | $1.20 | Free |
| LongCat-2.0 (standard) | $0.75 | $2.95 | Free |
| GPT-5.5 | $5.00 | $30.00 | Paid |
You can try LongCat-2.0 today through OpenRouter (the model formerly listed as Owl Alpha) and through Meituan’s own LongCat platform. The Hugging Face model card and GitHub repo are live for the eventual weight release, but at launch they note the open weights are still on the way. Full launch details are in Meituan’s official LongCat-2.0 announcement.
If you would rather not wire up an API key to test it, a multi-model assistant like Fello AI lets you swipe across frontier models from one chat, so you can pit an open-weight coder against the closed flagships on your own prompts and see which actually wins for your work.
Should You Use It?
For agentic coding, repository-scale work, and any task that benefits from a huge context window at a low price, LongCat-2.0 is worth testing now, especially while the promotional pricing and free cache hits are live. The real-world OpenRouter usage gives it more credibility than a typical launch, and the MIT license keeps the door open for commercial deployment.
Just keep expectations calibrated. The benchmarks are self-reported, the GPT-5.5 win is a one-test photo finish, and the open weights are not actually downloadable yet. Run it on your own code before declaring it a daily driver.
Conclusion
LongCat-2.0 is one of the most interesting open releases of 2026, a 1.6T MoE coding model that earned its reputation in production before it was even named, built entirely on Chinese silicon and priced to undercut the US flagships by an order of magnitude. The smart next step is to try it on a real coding task through OpenRouter, compare the output against a model you already trust, and decide based on your results rather than the leaderboard.
FAQ
What is LongCat-2.0?
LongCat-2.0 is Meituan’s open-source Mixture-of-Experts model with 1.6 trillion total parameters, roughly 48 billion active per token, and a native 1-million-token context window. It is built for agentic coding and was released under the MIT license on June 30, 2026.
Is LongCat-2.0 really better than GPT-5.5?
Only on one vendor-reported benchmark. Meituan says LongCat-2.0 scores 59.5 on SWE-bench Pro versus GPT-5.5’s 58.6, a margin under a single point. The numbers are self-reported and not independently verified, so treat it as competitive rather than clearly superior.
What was the Owl Alpha model?
Owl Alpha was the anonymous codename for LongCat-2.0 while it ran on OpenRouter for about two months. It became one of the platform’s most-used models, handling roughly 10.1 trillion tokens a month before Meituan revealed it as LongCat-2.0.
Was LongCat-2.0 trained without Nvidia chips?
Yes, according to Meituan. The company says LongCat-2.0 was trained and served on a cluster of more than 50,000 Chinese-made ASICs with no Nvidia GPUs, making it the first trillion-parameter model built entirely on domestic compute.
How can I use LongCat-2.0?
You can access it now through OpenRouter (formerly listed as Owl Alpha) and Meituan’s LongCat platform. The Hugging Face and GitHub pages are live, but the downloadable open weights are still listed as coming soon at launch.
