Meituan LongCat-2.5 Trained on Chinese Chips

Meituan on Nov. 2 released LongCat-2.5, a 1.6-trillion-parameter multimodal model built for long-running AI agent work such as sustained coding. The Chinese delivery and local-services giant says the model was trained entirely on domestic computing hardware and scores between Anthropic's Claude Opus 4.8 and Claude Opus 5 on several mainstream benchmarks.

Key takeaways

• LongCat-2.5 has 1.6 trillion total parameters, with about 48 billion active per token, according to TechNode.
• Meituan says training ran entirely on domestic Chinese compute, but it does not name the chips.
• The model is live in Meituan's chat app and API. Open weights on GitHub and Hugging Face are planned but not yet out.

What is Meituan LongCat-2.5?

Meituan's announcement describes LongCat-2.5 as a native multimodal model for "long-horizon" agent tasks. Those are jobs where an AI agent must work through many steps without losing track of the goal.

TechNode, citing IT Home, reported that the model handles images and text in a single token framework. Supported uses include visual coding, creative design and 3D tasks. Meituan says it expanded training on long task sequences, ran large-scale reinforcement learning on complex codebases and turned failures seen in real agents into repeatable training tasks.

LongCat-2.5 benchmarks vs. Claude Opus

Meituan says LongCat-2.5 lands between Claude Opus 4.8 and Claude Opus 5 on several mainstream benchmarks and beats Opus 5 on some. The announcement gives no per-benchmark scores, and no independent results had been published as of Nov. 3.

The company did publish efficiency figures. It says a 4-bit weight-and-activation (W4A4) version performs about as well as the unquantized model on the tasks it evaluated. Against an 8-bit (W8A8 INT8) baseline in an online stress test, Meituan reports about 40% higher throughput, about 16% less wait for the first output and about 20% less time per later token. It also cites an 86% gain in single-card throughput for long-context fine-tuning.

Context: China's push for "domestic model, domestic chip"

LongCat-2.5 follows LongCat-2.0, released in June. Meituan calls that model the first trillion-parameter model trained and served end to end on domestic compute. Its Hugging Face model card lists an MIT license and pretraining on more than 35 trillion tokens. It also reports an in-house SWE-bench Pro score of 59.5.

Caixin reported on Nov. 3 that the release shows Chinese tech firms leaning more on homegrown chips. It also comes as Chinese labs ship open models quickly, as when Qwen3.8 Max climbed above every American lab but two on one independent index.

Meituan is aiming the model at its own business rather than at selling raw compute. Its Miaoshou merchant workbench runs on the LongCat series and serves chains including Gu Ming and Yi He Tang. At its Aug. 28 earnings call, CEO Wang Xing said Meituan would not become a "token factory."

What's next

The open-weight release is the main date to watch. Once the weights appear on GitHub and Hugging Face, outside researchers can test Meituan's claims against Claude Opus on their own hardware.

FAQ

Is LongCat-2.5 open source?

Not yet. TechNode reports that GitHub and Hugging Face releases are planned. The earlier LongCat-2.0 weights are available under the MIT license.

How can developers try LongCat-2.5?

Meituan says the model can be tried at longcat.ai and called through an API launched the same day.

Which chips trained LongCat-2.5?

Meituan says only that training ran entirely on domestic compute. It has not named a chip vendor.

References

• 美团发布LongCat-2.5大模型，"国模国芯"实现新突破 — Meituan (Nov. 2, 2026)
• meituan-longcat/LongCat-2.0 — Hugging Face
• Meituan launches LongCat-2.5 multimodal AI model for extended coding tasks — TechNode (Nov. 2, 2026)
• Meituan Unveils New AI Model Trained on Homegrown Chips — Caixin Global (Nov. 3, 2026)