Meituan Open-Sources LongCat-2.0: 1.6 Trillion Parameters, Trained Entirely on Chinese Chips, Beats GPT-5.5 on SWE-Bench

A food-delivery company trained a model that outscores GPT-5.5 — on purely domestic Chinese chips — then gave it away for free.

Three Key Takeaways

On June 30, Meituan officially open-sourced LongCat-2.0 under the MIT license: 1.6 trillion parameters, free for commercial use. The architecture is a sparse Mixture-of-Experts (MoE) system that activates roughly 48 billion parameters per token (ranging dynamically from 33B to 56B), with a native 1-million-token context window. Pre-training consumed over 35 trillion tokens.

The entire training run used 50,000+ domestic Chinese ASIC accelerators — zero Nvidia GPUs. Meituan reports no rollbacks and no irrecoverable loss spikes across the full run. Before the open-source reveal, the model ran anonymously on OpenRouter under the alias "Owl Alpha" for two months, processing approximately 10.1 trillion tokens per month and topping the Hermes Agent leaderboard.

SWE-Bench Pro score: 59.5, beating GPT-5.5's 58.6. It also scored 70.8 on Terminal-Bench 2.1 and 77.3 on SWE-Bench Multilingual. API pricing is set at $0.75/M input tokens and $2.95/M output, dropping to $0.30/$1.20 during launch promotion with free cached reads. Full model weights have not yet been uploaded — both GitHub and Hugging Face pages show "coming soon."

WangDou's Take

The most devastating part isn't the parameter count — it's the two-month undercover operation. "Owl Alpha" sat on OpenRouter, developers happily used it for coding tasks, nobody suspected domestic hardware. Then the mask comes off: surprise, that model you've been shipping code with was trained on 50,000 Chinese ASICs. The argument is already over before the debate starts. SWE-Bench Pro 59.5 vs GPT-5.5's 58.6 is a 0.9-point gap, but the narrative damage is nuclear-grade: the Nvidia export-ban moat is being filled in, one shovel at a time, by the engineering team of a company best known for delivering lunch.

Source: VentureBeat · Decrypt · Cryptopolitan