LongCat-2.0 (Meituan)

Learning Objectives

• Understand what LongCat-2.0 is and how Meituan, a delivery company, became a frontier model lab.
• Learn why training a 1.6 trillion-parameter model entirely on domestic Chinese chips is a significant milestone.
• Evaluate when LongCat-2.0 is a sensible choice for agentic coding or open-weights deployment.

What Is LongCat-2.0?

LongCat-2.0 is the flagship open-source AI model from Meituan, the Chinese local-services and food-delivery giant. Released on June 30, 2026, it is a mixture-of-experts (MoE) model with 1.6 trillion total parameters, of which roughly 48 billion are active per token (the routing layer selects between about 33 billion and 56 billion depending on the input), paired with a 1 million-token context window.

The headline is not just the size — it is where the model was built. LongCat-2.0 is, by Meituan's account, the first trillion-parameter model trained and run entirely on domestically produced Chinese chips, on a cluster of fifty thousand AI accelerator cards. For an AI Pro Playbook reader, that makes it one of the clearest signals yet that China can now train frontier-scale models without Western silicon.

Why a Delivery Company Built a Frontier Model

Meituan is best known for moving food and packages across Chinese cities, not for AI research. But the same problems that make delivery hard — real-time routing, demand forecasting, dispatch across millions of orders — pushed Meituan to invest heavily in machine learning, and the company has also backed frontier Chinese labs, including a lead role in Moonshot AI's funding round. LongCat is Meituan's move from AI customer to AI builder.

The model is purpose-built for agentic coding — long, multi-step software tasks where the model plans, edits across files, runs commands, and iterates. That focus, plus the open-weights release, has made LongCat-2.0 a fast climber: in the days after launch it has been leading OpenRouter, the marketplace that routes developer traffic across models, which is a strong real-world adoption signal rather than a lab benchmark.

Architecture & Training

LongCat-2.0's significance rests on three facts:

• Mixture-of-experts efficiency. With 1.6 trillion total parameters but only about 48 billion active per token, inference cost and latency stay closer to a far smaller dense model while total knowledge capacity stays large — the same economics that make other open MoE models practical to self-host.
• Domestic-chip training. Earlier Chinese flagships such as DeepSeek's V4-pro used domestic chips only for inference; LongCat-2.0 used them for both the compute-intensive pre-training phase and inference, across more than 35 trillion tokens on a 50,000-card cluster of AI ASIC superpods.
• Open weights. The model and its resources are published on Hugging Face under the meituan-longcat organization, so developers can download, self-host, and fine-tune it.

Benchmark Performance

Meituan positions LongCat-2.0 as a near-frontier coding model, with its standout result on SWE-bench Pro, where it edges out GPT-5.5:

A single benchmark lead does not make LongCat-2.0 the best model overall — the closed US frontier models still lead on the broadest capability suites — but matching or beating GPT-5.5 on a respected coding benchmark, with open weights and on domestic hardware, is a genuine milestone.

How You Access It

Strengths

• Near-frontier coding — scores 59.5 on SWE-bench Pro, edging out GPT-5.5, and has been leading OpenRouter shortly after launch.
• Open weights — downloadable and self-hostable from Hugging Face, so teams can run it in their own environment and fine-tune it.
• Efficient MoE design — 1.6 trillion total parameters but only about 48 billion active per token keeps inference cost manageable.
• Long context — a 1 million-token window suits large codebases and long agentic sessions.
• Built on domestic chips — a milestone for training frontier models without Western accelerators.

Limitations & Considerations

• Coding-focused — the model is optimized for agentic software tasks; it is not pitched as a general-purpose assistant the way the big US chatbots are.
• Data and jurisdiction — using the hosted LongCat Chat or a Meituan-operated endpoint routes prompts to Chinese servers, subject to Chinese data law; self-host the open weights to keep prompts in your own jurisdiction.
• Young ecosystem — far fewer English-language tutorials, integrations, and community resources than ChatGPT, Claude, or Gemini.
• Self-hosting is heavy — running a 1.6 trillion-parameter MoE model locally still requires substantial GPU resources despite the low active-parameter count.
• Benchmarks need validation — a single coding-benchmark lead is a launch-day claim; verify on your own workloads.

Best Use Cases

Getting Started

1. Try the model for free in your browser at longcat.chat.
2. To self-host, download the open weights from Hugging Face.
3. Point your existing agentic-coding workflow at the model and run it against a few of your own real tasks.
4. Compare its output and cost to the model you use today before adopting it for production work.
5. For sensitive data, prefer the self-hosted open weights over the Meituan-hosted chat endpoint.

Key Takeaways

• LongCat-2.0 is Meituan's open-source flagship — a 1.6 trillion-parameter mixture-of-experts model for agentic coding, with about 48 billion active parameters per token and a 1 million-token context window.
• It scores 59.5 on SWE-bench Pro, edging out GPT-5.5, and has been leading OpenRouter in the days after its June 30, 2026 release.
• It is, by Meituan's account, the first trillion-parameter model trained and run entirely on domestically produced Chinese chips — a milestone in China's push to build frontier AI without Western silicon.
• The open weights are on Hugging Face, so the model can be downloaded, self-hosted, and fine-tuned — but the hosted endpoint routes prompts to Chinese servers, so self-host for sensitive data.
• A delivery company building a near-frontier model underlines how widely frontier AI capability is now spreading beyond the original handful of labs.