Meituan Drops 1.6T-Param Cat Out of the Bag, No Nvidia Required 🐱

Meituan officially unveiled LongCat-2.0 on June 30, a 1.6-trillion-parameter open-source mixture-of-experts large language model that the Beijing-based food delivery company says is the first trillion-parameter model to complete both training and inference on domestic Chinese hardware. The release confirms that the system spent roughly two months running anonymously on OpenRouter under the alias Owl Alpha, where it reached first place on the Hermes Agent workspace, second on Claude Code, and third across OpenClaw deployments by monthly call volume before its identity was disclosed.

LongCat-2.0 activates roughly 48 billion of its parameters per token, with that figure swinging between 33 billion and 56 billion depending on query complexity, and supports a 1-million-token context window. The pretraining run spanned more than 35 trillion tokens across a cluster of over 50,000 domestically produced accelerators, and Meituan said the process finished with "no rollbacks or irrecoverable loss spikes." The company added that the cluster was built around large-scale ASIC superpods, with chip-to-chip coordination managed through Huawei's Collective Communication Library (HCCL), a setup that mirrors how Nvidia's NCCL coordinates its own GPU clusters. The model architecture introduces LongCat Sparse Attention (LSA), described as a mechanism to scale efficiently to 1M-context tokens.

Pricing is central to the launch. Standard API access runs $0.75 per million input tokens and $2.95 per million output, cut to $0.30/$1.20 during the current launch promo, with cached context reads free of charge. That undercuts OpenAI's GPT-5.5 at $5/$30 per million tokens and Claude Sonnet 5's introductory $2/$10 rate, and lands close to DeepSeek V4-Pro's permanent $0.435/$0.87 and Xiaomi's MiMo-V2.5 Pro, which matched that same rate after its own May price cuts. Meituan also offers a token plan with packs of 1 billion tokens priced at around $60. DeepSeek's V4-Pro, by comparison, used Huawei chips only for inference while pretraining ran on Nvidia hardware.

LongCat-2.0 outperformed Google's older Gemini 3.1 Pro on Terminal-Bench 2.1 and SWE-Bench Pro, but still trails global frontier systems including OpenAI's GPT-5.5 and Anthropic's Opus 4.8 across the most demanding agentic and reasoning tasks. Lehigh University researcher Hanchi Sun called it the first model ever trained to near-frontier performance on 50,000 Chinese domestic accelerators, while tech analyst TP Huang said the launch puts to rest concerns about Huawei's Atlas-950 SuperPoDs. On X, analyst Yuchen Jin wrote, "It reminds me of Jensen Huang's point on the Dwarkesh podcast: export controls on Nvidia GPUs won't stop China. They'll just accelerate the development of AI that runs on Chinese chips."