According to monitoring by Dongcha Beating, Meituan has officially open-sourced the trillion-parameter model LongCat-2.0, which has a total of 1.6 trillion parameters and an average activation of about 48 billion, specifically designed for real Agentic Coding tasks. The architecture innovatively introduces LongCat sparse attention and N-gram embedding; the former reduces fragmented memory access through fluency perception indexing and hierarchical indexing, accelerating training and inference for long contexts of millions of tokens. The latter invests 135 billion parameters into the embedding layer while maintaining a sparsity of nearly 97% in the MoE, balancing parameter efficiency and structural stability. Post-training employs multi-teacher online distillation, categorizing experts into three types: Agent, Inference, and Interaction, seamlessly integrating through the MOPD architecture on domestic computing clusters. As the first trillion-parameter model to complete inference on a 50,000-card domestic computing cluster, LongCat-2.0 validates the mature capability of domestic chips to handle complex large model tasks. To address the multiple limitations of domestic chips in terms of memory, bandwidth, and interconnectivity, Meituan has made breakthroughs in three areas: model, chip adaptation, and deployment. At the model level, ScMoE utilizes the core control capabilities of domestic chips to achieve physical core-level parallelism between Dense and MoE branches, combined with KV-cache segmentation to alleviate the memory pressure of ultra-long contexts. At the chip adaptation level, Super Kernel reduces operator startup overhead, and Weight Prefetch hides I/O latency, maximizing hardware utilization under constrained conditions. At the deployment level, PD separation balances TTFT and TPOT, along with asynchronous Expert-Parallel load balancing to address uneven loads under high EP degrees. This open-source release also provides multi-precision versions such as BF16, FP8, and INT8, and aims to fully open the inference results optimized for domestic computing power, targeting smooth deployment of trillion-parameter model inference services even on existing and older domestic cards.

This page may contain third-party content, which is provided for information purposes only (not representations/warranties) and should not be considered as an endorsement of its views by Gate, nor as financial or professional advice. See Disclaimer for details.

Reward
like
Comment
Repost
Share

Comment

Add a comment

No comments

Trending Topics
View More
#
GTBurns2.57MInQ2
171.73K Popularity
#
PredictWorldCup🇦🇷vs🇪🇬
261.25K Popularity
#
VitalikUnveilsLeanEthereum
174.32K Popularity
#
gStocksTokenizedStocksLive
4.88M Popularity
#
BitcoinWhalesAdd270KInTwoWeeks
87.07K Popularity

Pinned

Sitemap

Meituan Open Sources Trillion-Parameter Model LongCat-2.0, Releases Inference Code for Domestic Chinese Chips

Trending Topics

GTBurns2.57MInQ2

PredictWorldCup🇦🇷vs🇪🇬

VitalikUnveilsLeanEthereum

gStocksTokenizedStocksLive

BitcoinWhalesAdd270KInTwoWeeks

Pinned