← Back to feed

Meituan, the Chinese food delivery and consumer tech giant, open-sourced LongCat-2.0 today: a 1.6 trillion total parameter MoE with 48 billion active parameters per token, native 1M-token context, and benchmark results that place it at the frontier tier for coding and agentic tasks.

The model narrowly beats GPT-5.5 on SWE-bench Pro (59.5 vs 58.6) and posts 70.8 on Terminal-Bench 2.1. Both are self-reported but corroborated by third-party analysis. SWE-bench Multilingual scores 77.3; FORTE (general corporate workflow simulation) reaches 73.2.

LongCat-2.0 has been the top-3 model globally on OpenRouter by call volume for several weeks under a closed preview before today’s public release.

The Hardware Story

The more significant data point may be the training stack. LongCat-2.0 was pretrained on over 50,000 domestic AI ASIC superpods, with no Nvidia hardware anywhere in the pipeline. Pretraining consumed more than 35 trillion tokens across more than a million accelerator-days, with no rollbacks or irrecoverable loss spikes.

That matters beyond the chip export control narrative. It is the first public demonstration that a frontier-scale MoE model can be trained end-to-end on non-Nvidia hardware with production-grade reliability. The infrastructure choices the team made (6D parallelism including a new EMBP dimension for N-gram Embeddings, Muon optimizer at scale, fault-tolerant monitoring with zero-manual-intervention link recovery) are described in the release paper and represent a serious engineering contribution independent of the model’s capabilities.

Architecture

LongCat Sparse Attention (LSA) is the headline architectural innovation. It extends DeepSeek’s Sparse Attention with three orthogonal efficiency improvements:

• Streaming-aware Indexing: Turns fragmented memory access into sequential reads, achieving coalesced HBM bandwidth
• Cross-Layer Indexing: One indexing pass serves several consecutive attention layers, reducing redundant compute
• Hierarchical Indexing: Two-stage coarse-to-fine scoring shrinks the candidate space the indexer processes per query

The result is efficient 1M-token native context without the quality compromises that typically accompany aggressive sparse attention at this scale.

A secondary innovation is N-gram Embedding: a 135B parameter module that expands the embedding space roughly 100x through N-gram token combinations. It improves token-level representations at a cost in VRAM that is far more efficient than adding equivalent-scale standard experts.

Post-Training Structure

LongCat-2.0 uses a MOPD (Multi-Output Post-training Distribution) architecture to fuse three independently trained expert groups:

• Agent Experts: Autonomous task execution, tool invocation precision, self-correction in agentic loops
• Reasoning Experts: Mathematics, STEM, multi-hop reasoning
• Interaction Experts: Instruction following, hallucination suppression, safety calibration

The model has been integrated into Claude Code, OpenClaw, and Hermes harnesses at the release date, and the team reports strong performance on repository-level edits and automated task execution in those environments.

Context in the Frontier Landscape

LongCat-2.0 does not displace Fable 5, Opus 4.8, or GPT-5.5 on SWE-bench Verified, where those models post results above 88-95%. SWE-bench Pro and SWE-bench Verified measure different things: Pro is harder, uncontaminated, and more representative of production software complexity.

The relevant comparison is that an open-source model from a food delivery company, trained on hardware that cannot be legally sold to China by Nvidia, just posted a higher SWE-bench Pro score than GPT-5.5. That closes an argument that was being made as recently as three months ago: that the frontier gap between proprietary US models and Chinese open-source alternatives was insurmountable.

It is not insurmountable. The gap on SWE-bench Verified remains large. The gap on agentic coding is now effectively zero for the models that matter most in production.