Meituan, China's food delivery and on-demand services giant, on Tuesday, 30 June 2026, open-sourced LongCat-2.0 — a 1.6-trillion-parameter large language model (LLM) that the company claims is the country's largest AI model trained entirely on home-grown semiconductor hardware. The Beijing-based company says the release marks a pivotal shift in how China deploys domestic chips, moving beyond inference-only use to full-scale pre-training on local silicon.

What LongCat-2.0 actually is

LongCat-2.0 carries 1.6 trillion parameters and supports a context window of 1 million tokens, placing it on par with DeepSeek's current flagship model, V4-pro, which launched in April 2026. Meituan open-sourced the model, making its weights publicly available — a move that echoes the open-source strategy that propelled DeepSeek to global attention earlier this year.

According to the company, LongCat-2.0 is the industry's first trillion-parameter model to complete full-process training and inference on a 50,000-card domestic computing power cluster, built entirely on what Meituan describes as 'large-scale clusters of tens of thousands of AI ASIC superpods.'

Why it matters: training, not just inference

The distinction between inference and pre-training is critical. Inference is the relatively lightweight process of running a finished model to answer queries. Pre-training — the phase where an AI model ingests massive datasets to learn foundational patterns — is orders of magnitude more computationally demanding, and has historically required high-end Nvidia GPUs.

According to Meituan, while DeepSeek-V4-pro relied on domestic chips only for inference, LongCat-2.0 used home-grown hardware for both inference and pre-training. If verified independently, this would represent a meaningful capability leap for China's domestic chip ecosystem — a sector under sustained pressure following successive rounds of US export controls targeting advanced semiconductors.

The competitive backdrop

The release positions Meituan — better known globally for food delivery than frontier AI — alongside dedicated AI labs such as DeepSeek, Zhipu AI, and OpenAI's rivals in China. The use of AI ASIC superpods — application-specific integrated circuits customised for AI workloads rather than general-purpose processors — points toward Huawei Technologies and its Atlas-950 SuperPods as the likely underlying hardware, though Meituan did not explicitly name the chip vendor in its announcement.

Independent benchmarking platform Artificial Analysis and other third-party evaluators have not yet published comparative assessments of LongCat-2.0's performance against global frontier models from Anthropic, Google, and OpenAI.

What's next

The open-sourcing of LongCat-2.0 invites scrutiny from the global research community, which will be the ultimate arbiter of whether the model's benchmark performance matches its architectural scale. For China's semiconductor industry, the more consequential question is whether the 50,000-card domestic cluster used in training can be replicated at commercial scale — and whether the Huawei Collective Communication Library and related software stack can sustain frontier-level training runs without the optimisation advantages embedded in Nvidia's mature CUDA ecosystem.

Meituan's next moves — whether it integrates LongCat-2.0 into its consumer super-app or licenses it to enterprise clients — will signal how seriously the company intends to compete in China's fast-consolidating AI platform market.