Meituan Releases Open Weights for LongCat 2.0 AI Model Under MIT License
Meituan has released the open weights for its LongCat 2.0 large language model under the MIT license. The model contains 1.8 trillion parameters, with approximately 48 billion active parameters used per inference step. The release provides developers with direct access to the weights, advancing a project first confirmed in recent training reports.
Earlier accounts indicated the company completed pretraining LongCat 2.0 on a cluster of Huawei Ascend 910 chips. The infrastructure is roughly ten times larger than the compute cluster used for DeepSeek’s V3 model at launch, processing an estimated 35 trillion training tokens. Technical disclosures note the model features 135 billion embedding parameters and reduces zero-communication experts from 256 to 128 compared to its predecessor.
From the sources (3 posts)
@teortaxestexif you think about it: 910С is roughly 90% of р800. So this is a cluster ≈10X larger than DeepSeek's at the time of training V3 (55 days), for maybe a 3x pretrain. Ok, mixed FP8 vs FP16, CANN… But I can't see it taking >5-6 weeks. Meitua
@eliebakouchlongcat 2.0 (1.6T, ~48B active) weights are now open under MIT license interesting design choice compared to previous iteration is that they keep the same attention shape and REDUCE the number of zero communication experts from 256 to 128
@multimodalartLe Chaton Long is here! The 1.8T behemoth of Meituan's LongCat-2.0 model is now open weights. A 1T+ parameter model trained entirely on chinese ASIC