"LongCat-2.0: Meituan's 1.6 Trillion Parameter MoE Model"

9:00 pm

Meituan just dropped something massive — and I mean that in the literal sense. Their new LongCat-2.0 model clocks in at 1.6 trillion total parameters with only 48 billion activated per token, making it one of the largest Mixture-of-Experts architectures ever publicly detailed.

What Makes It Interesting

The model was quietly powering the "Owl Alpha" endpoint on OpenRouter for a while before being identified as Meituan's work. During that time it racked up an impressive 11 trillion monthly token throughput — suggesting real-world production usage, not just a research paper stunt.

Key specs:

  • Architecture: Mixture-of-Experts (MoE)
  • Total parameters: 1.6 trillion
  • Active per token: ~48 billion (~3% activation ratio)
  • Context window: 1 million characters
  • Developer: Meituan (the Chinese food delivery / super-app giant)

The MoE Pattern Here

LongCat-2.0's 3% activation ratio puts it in the same ballpark as DeepSeek V4-Pro (3.1%, 49B active / 1.6T total) and Kimi K2 (3.2%). This is becoming the sweet spot for large-scale MoE: keep inference costs manageable by only activating a tiny fraction of the total parameter count per forward pass, while maintaining the knowledge capacity of the full model.

The 1 million character context window is also notable — it's competitive with Gemini 1.5 Pro and suggests LongCat-2.0 is designed for serious document analysis and multi-turn reasoning tasks.

Why It Matters Out of China

Chinese AI labs have been moving fast. LongCat-2.0 isn't a benchmark-chasing model (no public MMLU or HumanEval scores were shared), but the scale and the production deployment tell their own story. When a company like Meituan — which runs real-time logistics, payments, and food delivery at massive scale — deploys a 1.6T MoE model for internal use, it validates that these architectures work at internet scale. The open-source community should keep an eye on whether components of LongCat-2.0 get released, as Meituan has been more open with some of their earlier work.