LongCat-2.0 Open‑Sourced: Meituan Releases Domestic‑Chip Optimized Inference Code

Meituan announced the full open‑source release of LongCat-2.0, a trillion‑parameter language model (1.6 T total parameters, ~48 B active tokens) specifically engineered for Chinese domestic accelerator cards.

Three‑fold breakthrough:

Model level: Sparse attention (LSA) uses absorb computation, parallel Indexer & MLA prolog, and KV‑cache splitting to reduce I/O and memory pressure for ultra‑long contexts; ScMoE leverages chip‑level core control to run dense and MoE branches in parallel, cutting end‑to‑end latency and enabling efficient million‑token inference.

Chip‑adaptation level: A "Super Kernel" reduces operator count, while Weight Prefetch hides I/O latency; layer‑wise KV‑cache transfer over high‑speed inter‑chip links maximizes hardware utilization under limited memory and bandwidth.

Deployment strategy level: PD‑separated deployment balances prefill (TTFT) and decode (TPOT) by shrinking the expert‑parallel domain and splitting KV‑cache, combined with asynchronous expert‑parallel load balancing to avoid load imbalance at high expert counts. The stack supports constrained decoding, multi‑step scheduling and MTP optimizations.

Technical upgrades include:

LongCat Sparse Attention (LSA): Flow‑aware, cross‑layer, and hierarchical indexing reduce fragmented memory accesses and duplicate index calculations, accelerating million‑token training and inference while preserving model quality.

N‑gram Embedding: Added alongside MoE experts; with MoE sparsity near 97 %, allocating 135 B parameters to N‑gram embedding yields higher returns than expanding experts, while staying under 10 % of total parameters.

MOPD architecture: Post‑training multi‑teacher online distillation separates experts into Agent, inference, and interaction groups, endowing the model with execution, reasoning, and safe alignment capabilities.

The release provides BF16, FP8, and INT8 weight files, covering a range of precision needs, and includes inference code for both GPU (via SGLang PR) and NPU (via FluentLLM branch). Model weights, code repositories, and an API platform are linked on HuggingFace, GitHub, and ModelScope.

By open‑sourcing the model and the highly tuned inference stack, Meituan aims to enable reproducible deployment of trillion‑parameter models on legacy domestic cards, unlocking broader industrial value from existing compute resources.