clauding.de
DE
• 2 min read AI-generated

LongCat-2.0 Was Owl Alpha All Along — and Was Trained Entirely on Chinese Chips

Copy article as Markdown

Meituan revealed that the anonymous model 'Owl Alpha', which topped the OpenRouter charts for weeks, is actually LongCat-2.0 — a 1.6-trillion-parameter model trained without a single Nvidia chip.

Open Source China LLM LongCat Meituan
Featured image for "LongCat-2.0 Was Owl Alpha All Along — and Was Trained Entirely on Chinese Chips"

Imagine using an AI model for weeks that solves your coding tasks better than most alternatives. It has no name, no company behind it, just the codename ‘Owl Alpha’ on OpenRouter. Then it turns out it comes from China’s largest food delivery platform and was trained entirely on domestic Chinese chips.

The reveal

Meituan, China’s leading delivery service, released LongCat-2.0 on June 29 under an MIT license — and simultaneously revealed it was identical to ‘Owl Alpha’. Owl Alpha had topped OpenRouter’s developer usage rankings for weeks without anyone knowing where it came from.

The strategy was deliberate: Meituan wanted real developer feedback on production workloads before the name of a Chinese company created bias in the current geopolitical climate. It worked. By the time the origin was revealed, LongCat-2.0 already had an established track record with developers who had chosen it purely on merit.

The specs

LongCat-2.0 is a Mixture-of-Experts model with 1.6 trillion total parameters, activating an average of 48 billion per token. It has a native context window of one million tokens.

Benchmark results: 59.5% on SWE-bench Pro (narrowly above GPT-5.5 at 58.6%) and 70.8% on Terminal-Bench. That’s frontier-level performance — from a model under an MIT license.

Trained without Nvidia

The most geopolitically significant aspect: LongCat-2.0 was trained entirely on a cluster of 50,000 domestic Chinese ASICs. Not a single Nvidia H100. No A100s. No US export-controlled chips.

China promoted this as a proof of concept that its domestic chip ecosystem can train frontier-scale models. The Fable 5 export ban, which pulled Anthropic’s most capable model from the market for 19 days, accelerated exactly this kind of investment in chip-independent infrastructure.

What this means

For the open-source AI world, LongCat-2.0 is a milestone. MIT license means no regional restrictions, no usage prohibitions, full permission to fine-tune and redistribute. Together with GLM-5.2 and DeepSeek V4-Pro, there are now three credible open-weight alternatives to Western frontier models.

The Owl Alpha story also reveals something fundamental: when developers choose blindly based on quality, a model’s origin doesn’t matter. Until someone reveals the origin.

Sources: