MEITUAN DEBUTS CHINA’S BIGGEST AI MODEL TRAINED ON LOCAL CHIPS
- 1.6T parameters · MoE with ~48B active · 1M context
- Both the full training run and the large-scale deployment are built entirely on AI ASIC superpods.
- LongCat-2.0 is pre-trained on over 50K AI ASICs (No GPUs).
- it used the Huawei Collective Communication Library (HCCL) to improve training stability. HCCL is a chip-to-chip communication system similar to the Nvidia Collective Communication Library (NCCL).
- Competing models like DeepSeek-V4-pro relied on domestic chips only for inference, Meituan utilized domestic hardware for both inference and the highly intensive pre-training phase.
- Because the domestic accelerators possess less per-device memory than an Nvidia H800 (80 GB), Meituan address this challenge along two dimensions: parallelism strategy and memory management.
#Meituan #AI #SupremeCourtBlocksTrumpFromRemovingFedCook
- 1.6T parameters · MoE with ~48B active · 1M context
- Both the full training run and the large-scale deployment are built entirely on AI ASIC superpods.
- LongCat-2.0 is pre-trained on over 50K AI ASICs (No GPUs).
- it used the Huawei Collective Communication Library (HCCL) to improve training stability. HCCL is a chip-to-chip communication system similar to the Nvidia Collective Communication Library (NCCL).
- Competing models like DeepSeek-V4-pro relied on domestic chips only for inference, Meituan utilized domestic hardware for both inference and the highly intensive pre-training phase.
- Because the domestic accelerators possess less per-device memory than an Nvidia H800 (80 GB), Meituan address this challenge along two dimensions: parallelism strategy and memory management.
#Meituan #AI #SupremeCourtBlocksTrumpFromRemovingFedCook