Skip to content
Modelwire
Subscribe

Meituan's 1.6-trillion-parameter model trains without Nvidia chips

Source published ·Modelwire updated

Original coverage: The Decoder ↗·How Modelwire adds context

Illustration accompanying: Meituan releases a 1.6-trillion-parameter agent model that never touched an Nvidia GPU

The development

Meituan's release of LongCat-2.5 signals a meaningful shift in the geopolitics of large-scale model development. At 1.6 trillion parameters, the agent-focused model achieves competitive performance (88.4 on Terminal Bench, near Claude Opus 5 levels) while relying entirely on Huawei Ascend chips rather than Nvidia infrastructure. This demonstrates that non-US hardware ecosystems can now support frontier-scale training and deployment, reducing dependency on American semiconductor supply chains and validating alternative acceleration platforms for Chinese AI development.

Modelwire’s AI-generated summary of coverage from The Decoder.

Modelwire analysis

Analyst take

Our AI-generated reading of the wider context and the next developments to watch.

The more pointed detail is what LongCat-2.5 represents as a proof of concept for Huawei Ascend at scale: not just that it works, but that a 1.6-trillion-parameter run completed without Nvidia hardware at all, which is a different claim than 'we used some banned chips we smuggled in.'

This sits in direct tension with the forensic findings covered two days earlier, when a former OpenAI researcher identified Nvidia Blackwell fingerprints inside six Chinese models (The Decoder, Nov 13). That story suggested Chinese labs were still dependent on illicit Nvidia access. Meituan's release complicates that picture: the dependency may be real for some labs but is not universal. The Anthropic story from Nov 11 on CUDA portability is also relevant here, since it showed that chip-agnostic training is becoming more tractable generally, which lowers the barrier for Ascend to compete on real workloads rather than benchmarks designed to flatter it.

Watch whether a third-party lab reproduces LongCat-2.5's Terminal Bench score on a held-out task suite within the next 60 days. If the number holds under independent evaluation, Huawei Ascend has a credible claim as a frontier training platform; if Meituan declines to support external replication, the benchmark result deserves significant skepticism.

This interpretation is generated from the summary above and the archive coverage cited below. Our methodology · Report an error

Coverage behind this analysis

These archive entries ground the connection in our analysis. They are ordered by source publication date, with links to our coverage and the original sources.

  1. ·The Decoder

    Researcher finds Blackwell GPU signatures in Chinese AI models

    A former OpenAI researcher has published forensic evidence that Chinese AI labs including DeepSeek, Moonshot AI, MiniMax, and Ant Group are training models on Nvidia's Blackwell GPUs, circumventing US export restrictions. By analyzing characteristic signatures embedded in published model weights, the researcher identified at least five training runs representing over half a billion dollars in…

    Read Modelwire coverage →Original source ↗

MentionsMeituan · LongCat-2.5 · Huawei Ascend · Claude Opus 5 · Terminal Bench

MW

How this coverage is produced

Modelwire uses AI to generate summaries and context from source headlines, snippets, and selected archive coverage. Automated checks do not verify every claim, and items are not routinely reviewed by a person before publication. Zacaria Solis operates the site. Read the linked source for the full evidence and report errors through our corrections process.

Modelwire summarizes, we don’t republish. The Decoder originally reported this story as “Meituan releases a 1.6-trillion-parameter agent model that never touched an Nvidia GPU”. The full content lives on the-decoder.com. If you’re a publisher and want a different summarization policy for your work, see our takedown page.

More to explore

China trained frontier models on banned Nvidia chips, undermining export controls

Alibaba's Qwen4-Max cuts inference costs with sparse activation design

The Decoder·

Researcher finds Blackwell GPU signatures in Chinese AI models

The Decoder·
Meituan's 1.6-trillion-parameter model trains without Nvidia chips · Modelwire