Meituan trains 1.6T MoE model entirely without Nvidia GPUs

LongCat-2.0 is an open-source Mixture of Experts (MoE) language model developed by Meituan, scaling to 1.6 trillion total parameters with only around 48 billion parameters activated per token.

It is specifically tuned for coding agents and long-horizon tasks, but the real technical shocker lies under the hood: the entire model was pretrained and deployed without a single Nvidia GPU.

Instead, Meituan built the full training pipeline on AI ASIC superclusters, covering millions of accelerator-days and over 35 trillion tokens. Crucially, the team reported zero training rollbacks and zero irrecoverable loss spikes, signaling that massive frontier runs on alternative hardware are now production-ready. Industry watchers speculate the underlying silicon likely involves Huawei chips.

Under the hood

To manage long-context workloads without melting hardware, LongCat-2.0 introduces LongCat Sparse Attention (LSA).

LSA replaces random memory reads with streaming-aware indexing, turning fragmented high-bandwidth memory access into sequential reads. It also shares indexing across adjacent layers and uses a two-stage coarse-to-fine candidate search to shrink scoring overhead.

The model also bakes in 135 billion N-gram Embedding parameters, scaling memory-efficient parameters outside standard MoE routing.

Across internal benchmarks like Terminal-Bench 2.1, SWE-bench Pro, and FORTE, LongCat-2.0 trades punches with proprietary frontier models, closely trailing Claude Opus and GPT-5.5 while outperforming Gemini 3.1 Pro in several agent coding tests.

Hardware specs and how to run it

Meituan released the full weights under the permissive MIT license on Hugging Face, but running it locally is not for the faint of heart.

Even an FP8 quantized checkpoint exceeds 2TB on disk. You cannot squeeze this into consumer GPUs; serving it requires an enterprise multi-accelerator cluster.

On the software side, it works with standard frameworks like vLLM, SGLang, and Transformers, with native deployment scripts provided for both enterprise GPUs and NPUs. It integrates directly into developer harnesses like Claude Code, OpenClaw, and Hermes for automated repository editing and multi-step tasks. Anyone can test the web version at longcat.ai.

Why it matters

Unless you manage an enterprise GPU or NPU cluster, you will have to wait for the open-source community to drop aggressive 2-bit or 4-bit quantizations before running this locally.

Still, this release is a massive milestone for infrastructure independence. Nvidia has held a near-monopoly on frontier training because training runs on other hardware usually succumb to fatal stability and cluster communication crashes. Meituan just proved that a non-Nvidia ASIC stack can swallow 35 trillion tokens and produce a legitimate top-tier coding model, paving the way for wider adoption of alternative AI chips.