BeijingMeituan fit a 1.6-trillion-parameter model onto sub-80GB Chinese chips.

Domestic accelerators carry under 80GB of memory per card, constraining how large models train.

Meituan rebuilt parallelism around the supernode and moved the vision encoder out of the training graph, claiming 86% more throughput.

The benchmarks are Meituan’s own, including one it built, and the model trails Claude Opus 5 on Terminal Bench.

Hands-on tests found fast output and no data errors, and the model is open for outside testing.

How each outlet framed it
Zhidx
benchmarks LongCat-2.5 against Claude-Opus-5 and GPT-5.6-Sol, and reports a hands-on test of visual coding and an Amazon earnings deck

Sources: Zhidx