Meituan, China’s dominant food delivery and local services platform, announced on July 6, 2026 that it’s releasing LongCat-2.0 as an open-source model under the MIT license. The architecture is Mixture of Experts: Meituan reports 1.6 trillion total parameters with an average activation of 48 billion parameters per token, placing it in the same scale class as leading frontier MoE systems. Both parameter figures come from Meituan’s announcement via Tech in Asia; they haven’t been independently verified from readable source text.
The open-sourcing story has a wrinkle. Industry reporting indicates LongCat-2.0 had been operating under the name “Owl Alpha” on OpenRouter, a public model routing platform, before Meituan publicly claimed it on July 6. That means the model was serving real user traffic at scale before anyone knew Meituan had built it. The scale of that pre-announcement operation couldn’t be independently confirmed from available sources at publication time, so treat that framing as directional context rather than established fact.
The domestic chip training claim is the headline-grabbing element and the one with the weakest current source support. According to Meituan’s announcement, as reported by Caixin Global, the model was trained on a cluster of Chinese-made AI processors from vendors including Huawei, Moore Threads, and MetaX. Caixin’s article URL wasn’t accessible for independent verification, so the specific vendor list and cluster size can’t be confirmed as independently reported facts at this time. The significance of the claim, that a frontier-scale MoE trained end-to-end on domestic Chinese hardware, is worth tracking, but it warrants source recovery before treating it as established.
Disputed Claim
Why it matters
Self-reported benchmarks. Read carefully. Meituan claims LongCat-2.0 outperforms Claude Opus on instruction-following (IFEval) and mathematics (IMO-AnswerBench) evaluations. Those are vendor-evaluated comparisons with no independent corroboration, no Epoch AI evaluation has been published. The MIT license is straightforwardly significant: it means commercial use without restriction, which matters for teams evaluating open-weight alternatives. The pattern of Chinese labs releasing MIT-licensed frontier models, Zhipu AI’s GLM-5.1 among them, is worth noting as a deliberate ecosystem-building strategy, not just coincidence.
Context
LongCat-2.0 arrives in a period when Chinese open-source model releases are drawing serious practitioner attention. DeepSeek’s R-series demonstrated that non-NVIDIA training runs could produce competitive results; Xiaomi’s trillion-parameter agent model followed a similar reveal-after-deployment pattern. LongCat-2.0 fits that template, scale first, announce later, open-source for ecosystem reach.
What to watch
Don’t expect the performance claims to hold up unchallenged. Independent evaluation, particularly from Epoch AI or LMSYS, will either validate or qualify the IFEval and IMO-AnswerBench comparisons against Claude Opus. The domestic chip training claim needs source recovery: if Caixin’s reporting can be confirmed, it’s one of the more significant infrastructure stories of the year. Watch also for enterprise adoption signals given the MIT license, whether LongCat-2.0 shows up in OpenRouter routing volume is a concrete leading indicator of practitioner uptake.
What to Watch
TJS synthesis
Wait for independent benchmarks before drawing performance conclusions. The MIT license makes LongCat-2.0 worth downloading and testing against your specific workloads, that’s a low-friction evaluation path. The architecture and parameter scale are consistent with frontier MoE design; the chip training claim, if verified, would be more strategically significant than the benchmark comparisons. Bookmark this for re-evaluation when Epoch AI publishes or when readable source text for the Caixin reporting becomes available.
Sources: Nyu, VentureBeat.