Technology & Gadgets | July 2026 | 4 min read
⚡ Quick Takeaways
- Meituan has open-sourced LongCat-2.0, a 1.6-trillion-parameter Mixture-of-Experts AI model.
- The model supports a native 1-million-token context window for large codebases and documents.
- Meituan says LongCat-2.0 scored 59.5 on SWE-bench Pro, narrowly ahead of GPT-5.5’s 58.6.
- It was trained and deployed on more than 50,000 domestically produced Chinese AI chips.
- The MIT License and low API pricing make the model attractive for enterprise coding agents.
China’s AI race has gained a major new entrant after Meituan released LongCat-2.0, an open-source model built for agentic coding, long-context work and low-cost enterprise deployment.
| 1.6T total parameters in the MoE model |
1M tokens native context window for long inputs |
59.5 reported SWE-bench Pro score |
LongCat-2.0 Targets Coding Agents
Meituan has officially released LongCat-2.0 on GitHub, Hugging Face and its own platform, positioning the model as an open-source challenger to closed AI systems used for software engineering. The company says the release also confirms that LongCat-2.0 powered Owl Alpha, the anonymous model that had gained strong developer adoption on OpenRouter.
The model is not being pitched mainly as a chatbot. Meituan designed it for agentic coding, where an AI system must inspect a repository, use tools, make code changes, debug errors and complete connected tasks. That makes the launch especially relevant for developers and companies building automated software workflows.
LongCat-2.0 uses a 1.6-trillion-parameter Mixture-of-Experts architecture, but it activates only about 48 billion parameters on average for each token. The active range changes from roughly 33 billion to 56 billion, allowing the model to spend more compute on harder tasks while keeping simpler work lighter.
| Important Context: The headline benchmark lead is narrow and specific to SWE-bench Pro, so it should not be read as proof that LongCat-2.0 beats GPT-5.5 everywhere. |
Benchmark Lead Comes With Low Pricing
Meituan reports that LongCat-2.0 scored 59.5 on SWE-bench Pro, compared with GPT-5.5’s reported 58.6. It also lists strong results on Terminal-Bench 2.1, SWE-bench Multilingual and internal workflow-style evaluations. Reuters reported that Meituan claims the model matches or exceeds some leading proprietary systems on selected coding and agent benchmarks.
The pricing is also aggressive. Under standard billing, LongCat-2.0 costs $0.75 per million uncached input tokens and $2.95 per million output tokens. A limited-time promotion reduces this to $0.30 per million uncached input tokens and $1.20 per million output tokens. Context-cache hits are processed free of charge.
That free cache policy could matter for coding agents. Automated tools often need to repeatedly read the same repository, dependency files and documentation. If cached context does not create extra input charges, companies can run longer repository-inspection sessions at lower recurring cost.
| Model / Metric | Input Price | Output Price | Why It Matters |
| LongCat-2.0 promo | $0.30 / 1M | $1.20 / 1M | Lowest launch pricing tier |
| LongCat-2.0 standard | $0.75 / 1M | $2.95 / 1M | Commercial pay-as-you-go API |
| SWE-bench Pro | 59.5 | 58.6 GPT-5.5 | Narrow coding benchmark lead |
| Context cache | Free hits | Output billed | Cuts repeated repository costs |
| License | MIT | Open-weight use | Allows commercial integration |
Domestic Chips Add Strategic Weight
One of the biggest parts of the announcement is the training hardware. Meituan says LongCat-2.0 was trained and deployed on a cluster of more than 50,000 domestically produced Chinese AI chips. Reuters reported that the company presented this as evidence that large-scale AI training can now be done on Chinese compute clusters.
That matters because much of the global AI industry has relied heavily on Nvidia GPUs. If Chinese companies continue to train large models using domestic application-specific chips, it could reduce their dependence on US-controlled hardware supply chains and increase pressure on Western chipmakers.
The architecture also supports the long context window through LongCat Sparse Attention. Meituan says the system reduces the cost of attending across very large inputs by selecting important information more efficiently, helping the model handle full repositories, long documents and enterprise datasets.
| Hardware Signal: The launch is as much about China’s AI compute independence as it is about one model’s coding benchmark score. |
Open License Could Boost Enterprise Use
LongCat-2.0’s MIT License gives companies wide flexibility. Businesses can use, modify and integrate the model into proprietary products without being forced to release their private code. That makes it easier for software vendors, banks, healthcare companies and internal engineering teams to test private deployments.
For enterprises, the most useful scenario may be large-scale code migration. A development team could place a repository, logs and updated software development kit documentation inside the model’s context window. LongCat-2.0 could then identify dependencies, suggest changes, test failures and help prepare a pull request.
The model still needs independent testing outside company-provided charts. But its combination of open weights, long context, low pricing, domestic-chip training and strong coding scores gives developers another serious option in a market where access to premium closed models is becoming more expensive and more restricted.
| Enterprise Takeaway: LongCat-2.0 may appeal most to teams that need affordable coding automation, private deployment options and very large context windows. |
Disclaimer: This article is based on Meituan’s public release information and early external reporting. Benchmark claims should be treated as provisional until independently reproduced across more tests.
