Meituan LongCat-2.0 Release Shows Domestic Compute Can Handle Massive Models

Meituan released LongCat-2.0 on June 30 and open sourced the weights the same day. The model carries 1.6 trillion total parameters yet activates only about 48 billion during inference. It supports a native 1 million token context window and completed the full training and inference cycle on a cluster of 50,000 domestic accelerators.

The release puts direct pressure on other teams that still rely on overseas hardware for comparable scale. LongCat-2.0 reached SWE-bench Pro at 59.5 and SWE-bench Multilingual at 77.3, numbers that sit close to the current frontier while using only hardware available inside China.

What exactly changed

LongCat-2.0 uses four architectural pieces that had not been combined at this scale before. LSA sparse attention reduces the cost of the 1M context. Zero-compute experts skip unnecessary calculations inside the mixture-of-experts layers. ScMoE and MOPD fuse three separate expert groups labeled Agent, Reasoning, and Interaction. The result is a model that stays under 50 billion active parameters while still delivering agent-grade performance on software engineering benchmarks.

The training run finished entirely on domestic silicon. That fact alone removes one common objection against Chinese frontier models, namely that they depend on restricted hardware.

Why scale on domestic hardware matters now

Teams that build coding agents or long-document products have been waiting for a model that can hold a full codebase or long research corpus without chunking. LongCat-2.0 ships that capability today. Because the weights are open, any organization can run the model on whatever cluster it already controls instead of negotiating export-controlled GPUs.

Meituan also exposed a preview through OpenRouter and longcat.ai. Monthly call volume quickly rose into the global top three on OpenRouter. That traffic number gives an early signal that developers are already testing the model in production workflows.

Primary opponent in the current race

The clearest opponent line is between models that still require foreign accelerators and those that can train at trillion-parameter scale on domestic hardware. LongCat-2.0 lands on the domestic side. Every additional benchmark point it earns tightens the margin for teams that have not yet cleared that hardware hurdle.

Technical path that enabled the result

LSA sparse attention cuts the quadratic cost of attention over one million tokens. Zero-compute experts drop entire sub-networks when the router assigns them zero weight. ScMoE routes tokens across the three expert groups so that Agent, Reasoning, and Interaction skills remain specialized rather than averaged. MOPD then merges outputs from the three groups before the final layer. The combination keeps active parameter count low while preserving task performance.

These choices were validated end-to-end on the same 50,000-card cluster used for training. Inference latency and throughput numbers therefore reflect real cluster conditions rather than idealized lab runs.

Remaining uncertainties

SWE-bench numbers are strong, yet they do not yet cover every enterprise workload. Long-document retrieval and multi-turn agent sessions still need independent measurement. Meituan has released only limited ablation data, so it is not yet clear how much each new component contributes versus the sheer scale of the run. Independent labs will need time to reproduce the results on their own hardware.

What to watch next

Three signals will show whether LongCat-2.0 sets a durable precedent. First, how quickly other domestic labs publish similar training runs on the same cluster size. Second, whether OpenRouter call-share data remains above the top-five threshold after the initial novelty fades. Third, whether downstream agent frameworks adopt the 1M context window as a default rather than an experimental option.

If any of these three moves in the expected direction within the next quarter, the case that domestic clusters can support frontier training will strengthen. If none move, the result stays an interesting single-point achievement rather than a new baseline.