Meituan trains advanced LongCat-2.0 AI model with 1.6 trillion parameters using Chinese-made chips, excluding Nvidia GPUs for

Editorial illustration for Meituan trains 1.6 trillion-parameter LongCat-2.0 on Chinese chips, no Nvidia

Meituan trains 1.6 trillion-parameter LongCat-2.0 on...

Meituan trains 1.6 trillion-parameter LongCat-2.0 on Chinese chips, no Nvidia

By AI Daily Post Edited by Brian Petersen, Editor-in-Chief

June 30, 2026 • 2 min read

Meituan’s AI team has taken a bold step: a 1.6‑trillion‑parameter model, Long Cat‑2.0, trained entirely on Chinese hardware. The company says the effort proves “we now have the capability to train large‑scale models on domestic computing clusters.” To do it, Meituan ran the training on a farm of more than 50,000 home‑grown AI ASICs, processing over 35 trillion tokens. The Long Cat group, formed in 2023, shipped its first model just before the year ended, and the new version is already posting mixed results on public benchmarks.

It outpaces Gemini 3.1 Pro and GPT‑5.5 on SWE‑bench Pro (59.5) and SWE‑bench Multilingual (77.3), yet lags behind Claude Opus on the same tests. On IFEval, IMO‑AnswerBench and GPQA‑diamond it falls short of its Western peers. Meituan didn’t disclose the chip maker, and the model isn’t on HuggingFace, so independent checks are limited.

The rollout arrives amid U.S. export controls that have been in place since 2022, underscoring a growing push for home‑grown AI infrastructure.

Meituan's LongCat-2.0 shows China can train massive AI models without Nvidia Meituan trains a 1.6 trillion parameter AI model entirely on Chinese chips, no Nvidia required. "LongCat-2.0 has demonstrated that we now have the capability to train large-scale models on domestic computing clusters," the Chinese company said. Training ran on a cluster of more than 50,000 domestically made AI ASICs and covered over 35 trillion tokens.

The LongCat team has only existed since 2023. On some benchmarks, LongCat-2.0 beats leading Western models. On SWE-bench Pro (59.5) and SWE-bench Multilingual (77.3), it tops Gemini 3.1 Pro and GPT-5.5 but falls short of Claude Opus 4.7 and 4.8.

Meituan's LongCat-2.0 shows China can train massive AI models without Nvidia - THE DECODER

Why this matters

We now see a 1.6 trillion‑parameter model, LongCat‑2.0, trained entirely on more than 50,000 Chinese‑made AI ASICs, processing 35 trillion tokens without a single Nvidia chip. For developers accustomed to GPU‑centric pipelines, this demonstrates a viable, if still nascent, alternative hardware stack that could reduce reliance on foreign suppliers. Yet the article offers no benchmark data, leaving performance and cost efficiency open questions.

Founders may be tempted to explore domestic ASIC clusters, but scaling such infrastructure will likely demand significant capital and expertise that only a few firms currently possess. Researchers can study the training methodology, but without details on model quality or downstream applications, it’s unclear whether LongCat‑2.0 matches or exceeds comparable models built on more established platforms. The team’s rapid formation in 2023 and swift delivery of a massive model suggest strong internal momentum, but whether this momentum translates into a broader ecosystem of tools and talent remains uncertain.

In short, the achievement signals a potential shift in hardware sourcing, though practical implications for our community still need clearer evidence.