Main takeaway: Meituan released LongCat-2.0, a trillion-parameter model trained and run entirely on domestic chips. It has 1.6 trillion parameters, ~48 billion active per token, supports 1M token context, was pretrained on 3.5e14 tokens across 50k+ domestic chips, and is fully open-sourced under the MIT license with all weights and GPU/NPU inference code.
币界网消息，美团正式发布其首个从训练到推理均使用国产芯片的万亿参数大模型LongCat-2.0。该模型拥有1.6万亿总参数，单token激活参数约480亿，支持1m超长上下文。LongCat-2.0在超过5万张国产AI芯片集群上完成了35万亿token的预训练，成功验证了国产算力承载前沿大模型的工程稳定性。此外，LongCat-2.0宣布完全开源，采用MIT协议，放出了全部权重及完整的GPU和NPU推理部署代码。