Meituan Open-Sources LongCat-2.0: 1.6 Trillion Parameter Model Trained Entirely on Domestic Chinese Chips
Miles Bennett
Meituan open-sourced LongCat-2.0, a 1.6-trillion-parameter large language model trained end-to-end on Chinese-made chips — the first real-world validation of Huawei's Ascend ecosystem at trillion-parameter-scale pre-training.
What scale of model is LongCat-2.0?
1.6 trillion parameters with a 1-million-token context window — comparable to DeepSeek's flagship V4-pro released in April.
Meituan calls it China's first trillion-parameter model completed entirely on domestic chips, from pre-training through inference.
This means → the headline is less about model performance and more about which chips ran the full pipeline.
How does it differ from DeepSeek?
DeepSeek-V4-pro uses domestic chips only at the inference stage; pre-training still relied on other solutions.
LongCat-2.0 extends domestic hardware into pre-training — the most compute-intensive phase of AI development.
In plain terms = inference is "the model answering questions"; pre-training is "the model learning everything from scratch." The latter puts far greater stress on chips.
Why is pre-training so much harder?
Pre-training processes massive datasets to learn foundational patterns; it demands far higher communication stability and cluster scale than inference.
Meituan disclosed the model was built on a large-scale cluster of "tens of thousands of AI ASIC super-nodes."
This means → pre-training is not just a question of "are the chips fast enough" but "can tens of thousands of chips work together reliably."
Whose chips were actually used?
Meituan did not name its hardware supplier directly but disclosed the training used Huawei's HCCL (Huawei Collective Communication Library).
HCCL — Huawei's system for inter-chip communication — is the functional counterpart to Nvidia's NCCL.
This reflects that Huawei's Ascend-series chips are the underlying hardware; HCCL is exclusive to the Ascend ecosystem.
What does this mean for the market?
Huawei's AI chip ecosystem now has a real-world proof point at trillion-parameter pre-training, moving beyond inference-only substitution.
The core question remains open: whether domestic chips can reliably replace Nvidia GPUs across more frontier model training runs.
Put simply = Meituan proved "it can work"; whether "it can scale broadly" is a separate question the market is still watching.
Content is for reference only, not financial advice.