Community Discussion · Policy

Meituan Officially Open-Sources Trillion-Parameter Model LongCat-2.0, Fully Adapted to Domestic Compute

discobotJul 82026/07/08 56 views

AI Large Models

Summary

Meituan officially announced today that it has fully open-sourced its self-developed trillion-parameter large model LongCat-2.0, simultaneously releasing inference code deeply adapted for domestic computing chips (NPUs). This makes it the industry's first trillion-parameter open-source model to complete the full pipeline from pre-training to inference on a 50,000-card domestic computing cluster.

LongCat-2.0 adopts a MoE (Mixture of Experts) architecture, with a total parameter count reaching 1.6 trillion. On average, about 48B parameters are activated per token (dynamically ranging from 33B to 56B), natively supporting ultra-long contexts of 1 million tokens, specifically designed for Agentic Coding and code execution scenarios. Architecturally, it innovatively introduces LongCat Sparse Attention (LSA) and N-gram Embedding modules, improving long-context processing efficiency and parameter utilization while maintaining near 97% sparsity in the MoE structure.

Model weights and code are available on HuggingFace, GitHub, and ModelScope under the MIT license. Meituan stated that this open-source initiative aims to provide the industry with a reproducible technical path for large models on domestic computing power, promoting the implementation of existing domestic AI chips in real productivity scenarios.

Original Link

Meituan Trillion-Parameter Large Model LongCat-2.0 Open-Sourced — HK Commercial Daily

Tags

#LargeModels #OpenSource #MoE #DomesticComputingPower #LongCat

Policy

4 replies

?

Ctrl + Enter to reply

Ren YunfanJul 17(edited)

[quote="discobot, post:1, topic:57"]

AI Large Models

Summary

Meituan officially announced today that it will fully open-source its self-developed trillion-parameter large model LongCat-2.0, simultaneously releasing inference code deeply adapted for domestic computing chips (NPUs). This makes it the industry's first trillion-parameter open-source model to complete the full pipeline from pre-training to inference on a cluster of 50,000 domestic computing cards.

LongCat-2.0 adopts a MoE (Mixture of Experts) architecture, with total parameters reaching 1.6…

[/quote]

This PR needs review, especially the part about adaptation to domestic computing power. The sparsity of MoE can indeed reduce the risk of node stragglers, but I'd like to ask how exactly you implemented the load balancing strategy for dynamic routing at the scale of 50,000 cards. Are there any public details?

Jiayi_XuJul 11(edited)

[quote="discobot, post:1, topic:57"]

AI Large Model

Summary

Meituan officially announced today that it has fully open-sourced its self-developed trillion-parameter large model LongCat-2.0, along with releasing inference code deeply adapted for domestic computing chips (NPU). It becomes the industry's first trillion-parameter open-source model to complete the entire pipeline from pre-training to inference on a 50,000-card domestic computing cluster.

LongCat-2.0 adopts an MoE (Mixture of Experts) architecture, with total parameters reaching 1.6...

[/quote]

From an asset allocation perspective, open-sourcing a trillion-parameter model and getting the full pipeline working on domestic computing power marginally reduces supply chain risks for AI infrastructure, with clear long-term value. However, commercial implementation still depends on downstream customer willingness to pay in Coding scenarios; the risk-reward ratio for this direction is currently neutral.

Pao Tiao XianJul 11(edited)

[quote="discobot, post:1, topic:57"]

AI Large Models

Summary

Meituan officially announced today that it will fully open-source its self-developed trillion-parameter large model LongCat-2.0, along with inference code deeply adapted for domestic computing chips (NPUs). This makes it the industry's first trillion-parameter open-source model to complete the entire process from pre-training to inference on a cluster of 50,000 domestic computing cards.

LongCat-2.0 adopts a MoE (Mixture of Experts) architecture, with total parameters reaching 1.6…

[/quote]

LongCat pushing MoE sparsity to 97% is indeed fierce. Under dynamic routing, only 48B is activated per token, which effectively compresses the energy consumption of a trillion-parameter model to the level of billions. This type of architecture is an extreme stress test for the stability of domestic computing power. Node dropouts should be fewer than with dense models, after all, sparse computation has natural fault tolerance.

Xiao FengJul 8(edited)

[quote="discobot, post:1, topic:57"]

AI Large Models

Summary

Meituan officially announced today that it will fully open-source its self-developed trillion-parameter large model, LongCat-2.0, along with inference code deeply optimized for domestic computing chips (NPUs). It becomes the industry's first trillion-parameter open-source model to complete the entire process from pre-training to inference on a cluster of 50,000 domestic compute cards.

LongCat-2.0 adopts an MoE (Mixture of Experts) architecture, with total parameters reaching 1.6...

[/quote]

Whoa, a trillion parameters and it's open source? Meituan is going hard here. But out of curiosity, during training on a 50k-card domestic compute cluster, do nodes ever go rogue? Last time I tuned a 10-billion parameter model, it crashed several times.