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

Meituan has published the weights for LongCat-2.5, a model with 1.6 trillion parameters aimed at AI agent workloads. The company says it was trained entirely on Chinese hardware, and the sample deployment it describes runs on 194 Huawei Ascend chips. There is no mention of Nvidia GPUs anywhere in the release.
On Terminal Bench, a benchmark that tests an agent's ability to complete command-line tasks, LongCat-2.5 scores 88.4 points. That puts it just behind Claude Opus 5, according to the reported result.
The two notable claims here are scale and silicon. A 1.6-trillion-parameter model is large by any standard; models of that size are typically trained on clusters of thousands of accelerators, so fitting training onto a non-Nvidia stack is the part developers will watch. Whether the released weights are the full model or a variant suited to inference is not something the announcement settles, and the 194-chip figure describes a sample deployment rather than what training required.
Releasing weights also matters for teams that want to self-host: it opens the door to local fine-tuning and private agent deployments, though a model at this size implies substantial memory and multi-node serving in practice, which is my reading rather than a stated requirement.
Why it matters: Chinese labs are increasingly positioning domestic accelerators as viable for frontier-scale work, and if the Terminal Bench number holds up, agent builders get a credible open-weight alternative. For product teams, the practical questions are hosting cost and hardware compatibility, both inferences from the size and chip count rather than confirmed details.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Meituan has released the weights for LongCat-2.5, a 1.6-trillion-parameter model built for AI agents and trained entirely on Chinese hardware. On Terminal Bench, it scores 88.4 points, landing just behind Claude Opus 5. The sample deployment runs on 194 Huawei Ascend chips, and Meituan makes no mention of Nvidia GPUs. The article Meituan releases a 1.6-trillion-parameter agent model that never touched an Nvidia GPU appeared first on The Decoder .