Overview
- Meituan disclosed on June 30 that it open-sourced LongCat-2.0, a 1.6 trillion‑parameter mixture‑of‑experts model the company says is optimized for agentic coding tasks.
- The model uses a mixture‑of‑experts design with a one‑million‑token context window and dynamically activates roughly 33 billion to 56 billion parameters per token, which reduces compute per query compared with a dense model of the same size.
- Meituan says the model was pretrained and run end‑to‑end on a 50,000‑node cluster of domestically made AI accelerators and that training used Huawei’s HCCL to stabilize chip-to-chip communication, but the firm did not name the specific chip vendor and independent verification is incomplete.
- LongCat-2.0 already ran anonymously on the OpenRouter marketplace as 'Owl Alpha' and Meituan offers OpenAI‑ and Anthropic‑compatible API access with launch pricing that undercuts some rivals, while public weight availability on Hugging Face is reported but remains partly marked as 'coming soon'.
- If verified, the claim matters because training at this scale is far more compute intensive than inference and could weaken the leverage of U.S. export controls, lower costs for coders and small teams, and push competitors to speed up domestic chip and software development.