[Meituan Open-Sources Trillion-Parameter Model LongCat-2.0, Along with Inference Code Optimized for Chinese Domestic GPUs] According to monitoring by Dongcha Beating, Meituan has officially open-sourced its trillion-parameter model LongCat-2.0, featuring a total of 1.6T parameters with an average activation of approximately 48B. It is specifically designed for real-world Agentic Coding tasks. Architecturally, it introduces innovative LongCat Sparse Attention and N-gram Embedding. The former reduces fragmented memory access through flow-aware indexing and hierarchical indexing, accelerating training and inference for ultra-long contexts at the million-token scale. The latter allocates 135B parameters to the embedding layer, achieving a balance between parameter efficiency and structural stability, even with MoE sparsity approaching 97%. Post-training employs multi-teacher online distillation, categorizing experts into three types: Agent, reasoning, and interaction. These are seamlessly integrated on domestic computing clusters using the MOPD architecture. As the industry's first trillion-parameter model to complete inference on a 50,000-GPU domestic computing cluster, LongCat-2.0 validates the maturity of domestic chips in handling complex large-model tasks. To address the multiple constraints of memory, bandwidth, and interconnect on Chinese domestic chips, Meituan has made breakthroughs in three areas: model design, chip adaptation, and deployment. On the model side, ScMoE leverages the core control capabilities of domestic chips to achieve physical core-level parallelism between Dense and MoE branches, while KV-cache partitioning alleviates memory pressure for ultra-long contexts. On the chip adaptation side, Super Kernel reduces operator startup overhead, and Weight Prefetch hides I/O latency, maximizing hardware utilization under constrained conditions. On the deployment side, PD separation balances TTFT and TPOT, while asynchronous Expert-Parallel load balancing addresses uneven loads under high EP degrees. This open-source release includes multi-precision versions such as BF16, FP8, and INT8, and fully discloses inference optimizations tailored for domestic computing power. The goal is to enable smooth deployment of trillion-parameter model inference services even on existing and older domestic GPUs. --------------------------------- Click the link below to join the Dongcha Beating · Feishu AI News Channel for 24/7 global AI news and updates. [Original Link]