Meituan's LongCat-2.0: A 1.6T MoE Trained on AI ASICs, Not GPUs

Meituan released LongCat-2.0 on June 30, a 1.6-trillion-parameter mixture-of-experts model that does something no other model at its scale has done. It was trained entirely on AI ASICs — not Nvidia GPUs — on a 50,000-card domestic compute cluster. The pretraining spanned more than 35 trillion tokens across millions of accelerator-days with no rollbacks or irrecoverable loss spikes, according to the model card.

This is not a small experiment. LongCat-2.0 activates roughly 48 billion parameters per token, putting it in the same compute class as models like GPT-5.5 or Claude Opus 4.6. On SWE-bench Pro it scores 59.5, ahead of Gemini 3.1 Pro (54.2), GPT-5.5 (58.6), and Claude Opus 4.6 (57.3), per the model’s published benchmarks. It supports a native 1-million-token context through a new mechanism called LongCat Sparse Attention.

The implications for the AI industry are less about the benchmark scores and more about what the training run proves.

The ASIC question

For the past three years, the conventional wisdom has been that frontier-scale training requires Nvidia’s H100 and B200 GPUs, or at least their CUDA software stack. The export controls imposed by the U.S. government on advanced chips to China created a forced experiment: could frontier models be trained on alternative hardware at all?

LongCat-2.0 is the most credible answer yet. Meituan says the training cluster used domestic AI ASICs — application-specific integrated circuits designed for neural network workloads, not general-purpose GPUs. The company does not name the specific ASIC vendor, but the scale is unprecedented. A 50,000-card cluster running stable training across 35 trillion tokens is not a proof of concept. It is production infrastructure.

The model card emphasizes that the training run had “no rollbacks or irrecoverable loss spikes.” This is the kind of claim that only matters to people who have tried to train a large model on non-standard hardware. The CUDA ecosystem is mature; its debugging tools, profiling libraries, and distributed training frameworks have been refined over a decade. Building equivalent stability on a new chip architecture is a software engineering achievement that may be harder than the architecture itself.

Architecture as signal

LongCat-2.0 is a MoE model with 1.6 trillion total parameters and dynamic activation ranging from 33 billion to 56 billion per token. The architecture includes three innovations worth noting.

LongCat Sparse Attention reduces the quadratic complexity of full attention to linear, which is how the model handles 1 million tokens of context without memory collapse. The mechanism selectively attends to relevant tokens rather than all of them, similar in spirit to sparse attention methods from other labs but implemented for this specific hardware.

The “zero-computation experts” mechanism allows simple tokens to route through experts that require no compute at all, while complex tokens automatically receive more resources. This is a practical optimization for inference cost, especially on ASICs where every flop is more expensive relative to GPU alternatives.

The MOPD (Multi-Teacher On-Policy Distill) training pipeline starts from a LongCat SFT checkpoint, branches into three expert groups — Agent, Reasoning, and Interaction — and distills them back into one unified model. At inference time, a gating network dynamically routes each task to the most capable expert group rather than merging parameters. This is a distillation strategy designed to handle the fact that code generation, multi-hop reasoning, and instruction following benefit from different internal representations.

What this means for the AI supply chain

The most important sentence in the LongCat-2.0 release is not about benchmarks. It is this line from the model card: “demonstrating that we have the capability to conduct frontier-scale training on alternative hardware platforms.”

The U.S. export controls on advanced semiconductors were designed to slow Chinese progress on frontier AI by restricting access to Nvidia’s best chips. The assumption was that without H100s or B200s, Chinese labs could not train models at this scale. LongCat-2.0 does not disprove that assumption entirely — we do not know the performance characteristics of the ASICs used, or their cost per token — but it does show that the gap is bridgeable.

For the global AI industry, this raises a strategic question. If frontier models can be trained on ASICs rather than GPUs, the GPU supply chain becomes less of a bottleneck. Companies like OpenAI, Anthropic, and Google have been competing for Nvidia’s limited supply. If ASIC-based training becomes viable at scale, the competition shifts from hardware procurement to chip design and software optimization.

Meituan is not a chip company. It is a Chinese e-commerce and delivery platform that built an AI lab and trained a frontier model on custom hardware. That is a different kind of vertical integration than what the U.S. hyperscalers are doing with their own chips. It suggests that the next phase of the AI compute race may not be about who has the most GPUs, but who can design the most effective ASICs and the software stack to run them.

The open source question

LongCat-2.0 is released under the MIT license, per the Hugging Face model card. The weights are available for download, and the model has already accumulated 385 downloads and 140 likes on the platform as of July 5. A preview version has been ranking among the top three models by call volume on OpenRouter.

The open source release of a 1.6-trillion-parameter model trained on non-Nvidia hardware is a data point for the ongoing debate about open versus closed models. If LongCat-2.0 sees significant adoption in agentic coding workflows, it will validate the thesis that open models can compete with closed frontier models on practical tasks, not just on leaderboards.

The model’s performance on SWE-bench Pro and Terminal-Bench 2.1 suggests it is optimized for the kind of autonomous coding that tools like Claude Code and GitHub Copilot are targeting. The release includes detailed guides for agent workflow integration and self-hosting on smaller GPU setups, indicating that Meituan is targeting developers who want to run their own coding agents.

The open question

LongCat-2.0 does not specify the exact ASIC hardware used, the per-token training cost, or the inference efficiency relative to GPU-based models. The model card describes the training stability and the 1.5x MFU improvement from pipeline scheduling and memory optimization, but without a baseline comparison to Nvidia hardware, the numbers are hard to evaluate.

The real test will be adoption. If LongCat-2.0 gains traction among developers building agentic coding workflows, it will validate the ASIC-based training approach. If it remains a curiosity, the hardware advantage may not translate into practical capability.

For now, the model stands as the strongest evidence yet that frontier AI training is not permanently locked to Nvidia GPUs. That alone changes the geometry of the AI compute market.