Meituan's LongCat-2.0: First Trillion-Parameter Model Trained Entirely on Domestic Chinese Chips
The AMW Read
First trillion-parameter model trained on exclusively domestic Chinese chips resolves a key open debate about sovereign AI feasibility, with cross-segment implications for compute economics and geopolitics.
Meituan's LongCat-2.0: First Trillion-Parameter Model Trained Entirely on Domestic Chinese Chips
Chinese food-delivery and local-services giant Meituan has released LongCat-2.0, a 1.6-trillion-parameter mixture-of-experts (MoE) model that claims to be the world's first trillion-parameter large language model trained and deployed entirely on domestic Chinese AI chips—zero Nvidia hardware. The model uses a self-developed MoE architecture with approximately 48 billion activated parameters per token and supports a 1-million-token native context window. LongCat-2.0 was trained on a cluster of 50,000 domestic chips, and has already been validated in the open market: under the pseudonym "Owl Alpha" on OpenRouter, it became the top-called open-source model globally for agent developers over the past two months, surpassing Hermes, Claude Code, and OpenClaw in monthly API invocations.
Why it matters: LongCat-2.0 is a milestone for China's sovereign AI compute ecosystem. It demonstrates that a domestic chip stack can now support frontier-scale model training from scratch—previously a capability limited to Nvidia's CUDA ecosystem. This validates a strategic shift that Meituan began in 2023, accepting longer R&D cycles in exchange for insulation from US export controls. It also exemplifies the "context-engineering moat" pattern: LongCat-2.0's sparse attention mechanism (LongCat Sparse Attention, LSA) was redesigned for agent workloads with long contexts, and its competitive inference pricing—less than RMB 0.10 per query at scale—signals a deliberate cost-optimization strategy aimed at high-volume, high-throughput deployment scenarios.
Grounded expert take: LongCat-2.0 updates the foundation-model player map by confirming that a major Chinese internet platform—not a pure AI lab—can deliver frontier-scale training purely on domestic silicon. The structural signal is strongest for capital-cycle and compute-economics dynamics: Meituan's success reduces the perceived risk of non-Nvidia training infrastructure, potentially unlocking more domestic capital for alternative chip ecosystems. However, the model's strong agent-benchmark performance under an anonymous identity (Owl Alpha) suggests that quality parity with Nvidia-trained models is achievable, a finding that could reshape how hyperscalers and sovereign AI projects allocate compute budgets. The open debate about whether domestic Chinese chips can support frontier LLM training is now partially resolved—the answer is yes, with the caveat that engineering overhead remains higher than Nvidia's mature stack.