Meituan has released LongCat-2.0, a 1.6 trillion-parameter Mixture-of-Experts model with roughly 48 billion parameters active per token, and confirmed it is the system that powered “Owl Alpha,” a stealth model that had quietly climbed to top three globally by daily volume on OpenRouter. The reveal turns a months-long guessing game among developers into a straightforward product launch. It also gives Meituan, a company known to most outside China for food delivery, a legitimate claim to a frontier-scale coding model already proven in production traffic.
The stealth-to-reveal pattern is not new. Labs including Anthropic and Google have shipped unlabeled models onto public leaderboards before, using real usage data as a benchmark that marketing copy cannot fake. What differs here is the scale of adoption prior to disclosure: Meituan’s own account says Owl Alpha reached number one on Hermes Agent and number two on Claude Code by monthly volume, according to the company’s LongCat account on X. That is developer behavior, not a curated eval score, and it happened before anyone knew who built the model.
LongCat-2.0 targets agentic coding, tool use, and long-context work specifically. It ships with a 1 million-token context window, a maximum output length of 128,000 tokens, and a custom attention mechanism the company calls LongCat Sparse Attention, built to keep inference efficient at million-token scale. Meituan says pretraining ran across more than 35 trillion tokens on AI ASIC superpods, with hundreds of billions of additional tokens dedicated to long-context training data.
Access runs through the LongCat API Platform, with both OpenAI-compatible and Anthropic-compatible endpoints. That compatibility choice is a distribution decision: it lets LongCat-2.0 slot directly into Claude Code, OpenClaw, OpenCode, Kilo Code, and Codex-style tooling without developers rewriting integration code. Pricing is $0.75 per million uncached input tokens, $0.015 per million cached input tokens, and $2.95 per million output tokens, with limited-time discounts and 30-day token packs where cache hits do not count against the pack.
One qualifier matters for anyone evaluating this as an open-weight alternative: the GitHub repository is public under an MIT license, but the actual model weights are listed as forthcoming on both GitHub and Hugging Face. LongCat-2.0 is available today only as a hosted API product. Teams hoping to self-host or fine-tune the model will need to wait for a separate weights release, and Meituan has not given a date.
The company frames LongCat-2.0 as the flagship entry in a broader multimodal push that already includes LongCat-Flash-Chat, LongCat-Video, LongCat-Image, and LongCat-Next. That positions Meituan alongside Alibaba and ByteDance as a Chinese consumer platform company building frontier-scale general models rather than narrow, product-specific ones. The OpenRouter volume data suggests developers were already choosing the model on merit, without knowing its origin, which is a harder claim to dismiss than a self-reported benchmark table.
For teams currently routing agentic coding workloads through OpenRouter, the practical move is to check whether Owl Alpha traffic is already flowing through existing configurations and confirm pricing before the limited-time discounts expire. Anyone evaluating LongCat-2.0 against open-weight options should treat it as API-only until Meituan actually publishes weights.
Reported by TestingCatalog on June 30, 2026.
