Meituan LongCat-2.0: How a 1.6T Parameter Model Runs on 50,000 Domestic GPUs

In July 2026, Meituan open-sourced LongCat-2.0, a trillion-parameter model trained end-to-end on a 50,000-card domestic GPU cluster. It's the first model of this scale to complete both training and inference on Chinese-manufactured chips—a milestone that matters more than any benchmark score.

Model Architecture: 1.6T Total, 48B Active

LongCat-2.0 uses a Mixture-of-Experts architecture: 1.6 trillion total parameters with approximately 48 billion activated per token, dynamically ranging from 33B to 56B. The design philosophy is clear—massive capacity for knowledge coverage, sparse activation for inference efficiency.

Its target domain is Agentic Coding: not a chatbot that waits for prompts, but an agent that understands repositories, plans multi-step operations, executes code, and verifies results autonomously. To support this, the model natively handles 1 million token context windows—enough to ingest entire codebases in a single pass.

Three Architectural Innovations

LongCat Sparse Attention (LSA) addresses the quadratic complexity of attention over long contexts. Using flow-aware indexing, cross-layer indexing, and hierarchical indexing, LSA reduces fragmented memory access and redundant index computation while preserving model quality. The net effect: you only attend to what actually matters.

N-gram Embedding takes a different approach to parameter scaling. With MoE sparsity already at 97%, adding more experts yields diminishing returns. Instead, LongCat-2.0 invests 135B parameters in N-gram Embedding—an expansion path independent of Transformer layers, constrained to under 10% of total parameters. The return on this "investment" significantly outperforms further expert expansion.

ScMoE (Sparsity-Controlled MoE) leverages the fine-grained core-control capabilities of domestic chips to execute Dense and MoE branches in physical-core-level parallelism, squeezing down end-to-end latency. This is hardware-aware design at its best—tailoring architecture to chip characteristics rather than brute-forcing compute.

Engineering on Domestic Hardware

The real story of LongCat-2.0 is not the model itself but the engineering stack that makes it work. Domestic GPUs come with three constraints: limited VRAM, limited bandwidth, and limited interconnects. The team tackled these on three fronts:

• Model-level: Absorb computation patterns, parallel MLA prolog processing, and KV-cache sharding to reduce I/O and memory pressure during long-context inference.
• Chip-adaptation: Super Kernels to minimize operator count and launch overhead; Weight Prefetch to hide I/O latency behind computation.
• Deployment: Prefill-Decode separation with sequence parallelism for long inputs and KV-cache sharding for low per-card memory during decode.

The philosophy is "hardware-software co-design"—adapting the software stack to hardware constraints rather than waiting for newer silicon. The result: trillion-parameter inference on existing hardware, including older cards.

What's Actually Open-Sourced

This isn't a "weights-only" release. The full package includes:

• Multiple precisions: BF16, FP8, and INT8 versions covering everything from high-end GPUs to edge deployment.
• Inference code: GPU implementation submitted upstream to SGLang (PR #30042); NPU implementation available on GitHub.
• Weights: Available on HuggingFace, GitHub, and ModelScope simultaneously.

The strategy is transparent: Meituan wants the ecosystem to adopt domestic GPU inference stacks. By providing an end-to-end recipe, they're lowering the barrier for other teams to deploy trillion-parameter models on existing hardware.

Post-Training: Multi-Teacher Distillation

LongCat-2.0's post-training uses multi-teacher online distillation with three specialized expert categories:

• Agent experts: Autonomous execution and tool use
• Reasoning experts: Adaptive reasoning and deep thinking
• Interaction experts: Safety alignment and natural dialogue

These are fused through MOPD (Multi-Objective Policy Distillation) on the same domestic GPU cluster, producing a single model that combines deep reasoning, autonomous execution, and precise interaction.

Why This Matters

The timing is telling. In the first half of 2026, the AI industry narrative shifted decisively from "whose model is bigger" to "whose model is usable." At the AI Engineer World's Fair, "engineering" replaced "model racing" as the dominant theme.

LongCat-2.0's significance isn't another leaderboard entry—it's proof that domestic GPU ecosystems can sustain world-class AI workloads. For developers, a 48B-active model deployable on existing hardware is infinitely more valuable than a 600B model you can only read about.

That might be LongCat-2.0's most fundamental competitive advantage. It's not just open weights—it's an open invitation to build on hardware you already have.