美团发布LongCat-2.0，1.6万亿参数大模型完全基于国产芯片训练
阅读原文· the-decoder.com美团发布LongCat-2.0，参数规模达1.6万亿，完全基于超5万颗国产AI ASIC芯片集群训练，覆盖35万亿模型token。该模型在SWE-bench Pro（59.5）和SWE-bench Multilingual（77.3）上超越Gemini 3.1 Pro和GPT-5.5，但落后于Claude Opus 4.7和4.8。在IFEval（90.0）、IMO-AnswerBench（81.8）和GPQA-diamond（88.9）上则与Gemini、GPT-5.5存在差距。美团未透露具体芯片厂商，模型暂未在HuggingFace开放，独立验证困难。项目团队始于2023年，首个模型于去年底交付。
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. Its first model shipped late last year.
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. On other tests like IFEval (90.0), IMO-AnswerBench (81.8), and GPQA-diamond (88.9), it trails Gemini and GPT-5.5 by a wide margin in some cases.
The message to Washington is hard to miss. Despite US export controls in place since 2022, China appears to have produced its first competitive trillion-parameter model trained entirely on domestic hardware. Meituan didn't name the specific chip maker, though. And the model isn't yet available on HuggingFace, making independent verification difficult.