LongCat-2.0

Meituan's new flagship: a 1.6-trillion-parameter MoE activating ~48B per token, with a dedicated LongCat Sparse Attention mechanism and a 1M-token context (trained on hundreds of billions of tokens of 1M-context data). Built for coding and agentic, long-horizon tasks; open-sourced under MIT (weights "coming soon" at launch).

The headline is the hardware: LongCat-2.0 was both pre-trained and served entirely on a ~50,000-chip cluster of domestic Chinese "AI ASIC superpods" — no NVIDIA. Meituan did not name the chipmaker but said it used the Huawei Collective Communication Library (HCCL) to improve training stability (analysts point to Huawei Ascend or Cambricon). This makes it China's largest model pre-trained (not merely run) on domestic silicon — a step beyond DeepSeek-V4-pro, which used home-grown chips only for inference. Released June 30, 2026; Meituan describes its performance as comparable to Google's Gemini 3.1 Pro. Not yet scored by Artificial Analysis.