Meituan Drops 1.6T-Parameter LongCat-2.5, Built Without Nvidia GPUs
Meituan released the weights for LongCat-2.5 on November 15, 2026, a 1.6-trillion-parameter AI model aimed at agent workflows and built on Chinese-made accelerators instead of Nvidia GPUs. The deployment guide leans on 194 Huawei Ascend Atlas A2 chips, and the MIT-licensed weights let developers run it on their own hardware.
- Original source: lapaasvoice.com.
- The source published it on Nov 15, 2026 at 4:00 PM.
- No second outlet has reported it yet. Treat it as a single-source claim.
Meituan isn't teasing anymore. The Chinese tech giant dropped the model weights for LongCat-2.5 on November 15, 2026, roughly a week after first showing the thing off. The 1.6-trillion-parameter system was built and runs on Chinese-made AI accelerators — not a single Nvidia GPU involved, per Meituan and the model's technical documentation.
The architecture is mixture-of-experts: LongCat-2.5 holds 1.6 trillion parameters in total but only activates about 48 billion of them per token, with specialised groups of neural-network parameters handling different slices of a task. That keeps the model big on capacity while holding down the compute needed for each individual operation. It's built for autonomous agents rather than plain chatbot Q&A — planning and executing multi-step work like writing software, navigating development environments, image and video understanding, and analysing information.
The hardware story is the headline. Meituan's deployment documentation describes a configuration of 194 Huawei Ascend accelerators from the Atlas A2 series to run the model, split into groups by workload. There are no instructions for running this version on Nvidia GPUs — the guide sticks to Huawei's domestic accelerator infrastructure. Meituan says the predecessor, LongCat-2.0, was trained on a cluster of more than 50,000 Chinese-made AI chips, so the domestic-hardware strategy isn't a one-off.
Context: US export restrictions have limited Chinese tech companies' access to some advanced Nvidia processors, which is exactly why a trillion-parameter model skipping Team Green entirely lands with a thud. Nvidia GPUs have historically been central to building large AI systems worldwide. To be clear, this doesn't mean Chinese accelerators have replaced Nvidia everywhere — it shows Meituan built and shipped this particular model on a domestic ecosystem.
LongCat-2.5 comes out under the permissive MIT licence, so developers can download, modify and commercially use the weights, and run the system on their own infrastructure instead of depending on Meituan's hosted services. The LongCat AI research team inside Meituan is using the release to plant a flag in China's crowded AI race, where a complete deployment pathway — compatible hardware, software frameworks and optimisation tools — can matter as much as the model itself.
TopicsMeituan · LongCat-2.5 · LongCat-2.0 · Huawei · Nvidia · Atlas A2 · LongCat AI
Source: lapaasvoice.com
Written by AI from verified reporting. Not financial advice.

