Meituan LongCat-2.0: A 1.6T MoE model trained and deployed entirely on domestic chips

实测美团LongCat-2.0，国产芯片长出来的万亿大模型

Meituan released LongCat-2.0, a 1.6T total parameter MoE model activating ~48B per token, trained and deployed entirely on 50,000 domestic chips with over 35T tokens and no rollbacks or unrecoverable loss spikes. Agent performance stands out: it matches Gemini 3.1 Pro on Terminal-Bench 2.1 and SWE-bench Pro coding tasks, and ties Claude Opus 4.6 on FORTE general agent tasks. It offers up to 1M context and 128K max output, using LSA sparse attention and N-gram Embedding for long-context and tool-calling optimization. The API is live with OpenAI and Anthropic compatibility, ready for Claude Code and Codex workflows.

Why it matters: Meituan LongCat-2.0 is the first publicly disclosed trillion-scale MoE model trained end-to-end on domestic chips, matching Gemini 3.1 Pro on coding/agent tasks and Claude Opus 4.6 on general benchmarks. 50K domestic GPUs and 35T tokens without training collapse is itself a si...