Owl Alpha

Owl Alpha is the anonymous deployment of Meituan's LongCat-2.0 model on OpenRouter. It is a 1.6-trillion-parameter Mixture-of-Experts architecture, activating only about 48B parameters per token. The model incorporates LongCat Sparse Attention (LSA) to efficiently handle up to 1 million tokens of context, and uses a gating mechanism that routes tasks to three specialized expert groups: Agent, Reasoning, and Interaction. A key innovation is Zero-Compute Experts, which dynamically selects between 33B and 56B active parameters, avoiding wasted computation on irrelevant experts.

In benchmark evaluations, Owl Alpha scored 70.8 on Terminal-Bench 2.1 and 59.5 on SWE-bench Pro, close to leading closed-source models. During its stealth period on OpenRouter, it reached #1 by daily volume in the Hermes Agent category, and ranked #2 and #3 in Claude Code and OpenClaw tasks respectively, indicating strong real-world demand for agentic coding workloads.

However, the model was trained on tens of thousands of domestic AI accelerators (the company mentioned over 50K ASICs), making independent fine-tuning or self-hosting extremely resource-intensive.