Meituan Open-Sources LongCat-2.0, a 1.6-Trillion-Parameter Coding Model It Says Was Trained End-to-End on Chinese Chips
Meituan released LongCat-2.0, a 1.6T-parameter MoE model under an MIT license, claiming it is the first trillion-parameter model trained and served entirely on domestic ASICs.
Overview
Meituan, the Chinese food-delivery and local-services company, has open-sourced LongCat-2.0, a large language model with 1.6 trillion parameters and a native 1-million-token context window, according to the South China Morning Post. The company describes it as “the industry’s first trillion-parameter model to complete full-process training and inference on a 50,000-card domestic computing power cluster” — a claim that, if accurate, would make it the largest AI model to date trained end-to-end without Nvidia hardware, per the South China Morning Post.
The model was released on June 30, 2026 under an MIT license, and its weights are published on Hugging Face, according to the official Hugging Face model card.
What We Know
LongCat-2.0 is a sparse mixture-of-experts model with 1.6 trillion total parameters but only about 48 billion activated per token, according to the Hugging Face model card. The active count varies dynamically between roughly 33 billion and 56 billion per token, as reported by MarkTechPost.
To handle its 1-million-token context, Meituan introduced a linear-complexity attention mechanism it calls LongCat Sparse Attention (LSA), according to MarkTechPost. The Hugging Face model card describes LSA as combining three components: Streaming-aware Indexing, which “reshapes the token selection budget to combine hardware-aligned contiguous access with dynamic random selection”; Cross-Layer Indexing, which “leverages the empirical stability of attention saliency across adjacent layers”; and Hierarchical Indexing, which “uses a coarse-to-fine, two-stage scoring scheme.” The company reports the model was trained on more than 35 trillion tokens, per the Hugging Face card.
On company-reported benchmarks, LongCat-2.0 scored 59.5 on SWE-bench Pro, 70.8 on Terminal-Bench 2.1, 77.3 on SWE-bench Multilingual, and 79.9 on BrowseComp, according to the Hugging Face model card. MarkTechPost notes that the SWE-bench Pro figure edges out a reported 58.6 for GPT-5.5, but adds that the numbers are self-reported and have not been independently verified. The model is integrated with mainstream agent harnesses including Claude Code, OpenClaw, and Hermes, per the Hugging Face card, and Meituan lists API pricing of $0.75 per million input tokens and $2.95 per million output tokens, with a launch promotion of $0.30 and $1.20, according to MarkTechPost.
Before its identity was disclosed, the model ran anonymously on the OpenRouter marketplace under the codename “Owl Alpha,” according to NYU Shanghai RITS.
The Hardware Claim
The most consequential claim is about the silicon underneath. Meituan says LongCat-2.0 was trained and served on a cluster of tens of thousands of domestic AI ASIC superpods, with the South China Morning Post reporting a 50,000-card domestic computing power cluster. The distinction Meituan draws is between training and inference: the South China Morning Post notes that DeepSeek’s flagship V4-pro model, which launched in April, relied on home-grown chips only for inference, whereas LongCat-2.0 used domestic hardware for pre-training as well.
The release extends a trend The Machine Herald has tracked, following Zhipu AI’s GLM-5, which was billed as the first frontier model trained entirely on Chinese chips. LongCat-2.0 pushes the same claim to a larger scale and to the full training-plus-inference pipeline.
What We Don’t Know
Meituan has not publicly identified the specific vendor or model of the domestic ASICs used, beyond describing them as domestic superpods, in the coverage reviewed. The benchmark figures are self-reported and, as MarkTechPost notes, have not been independently reproduced, so LongCat-2.0’s standing relative to Western frontier models remains to be confirmed by third-party evaluation.