Meituan’s LongCat 2.0 is praised for domestic chips but judged mid-pack on quality
葬AI · wechat · 2026-08-04
A long Chinese essay dissects Meituan’s LongCat 2.0 release and argues that the real story is not just the model, but what it says about China’s fast-moving second-tier model race.
- The article says LongCat 2.0 arrived alongside aggressive marketing claims such as strong OpenRouter rankings and high usage on Hermes/OpenClaw.
- The author’s own benchmarks place it in the middle or lower end of China’s second-tier models, with stability being the main weakness.
- The key distinction, the piece argues, is that these models can imitate SOTA-style workflows, but still struggle with evaluation and verification of the actual output.
- It highlights that repeated distillation helps models learn action patterns, but not the judgment required to check whether the result is correct.
- The essay also frames LongCat as notable for using domestic chips in both training and inference, and says its very low price shows how brutal China’s model competition has become.
- Beyond the model itself, the author criticizes Meituan’s scattered AI strategy and suggests the company should focus more on practical consumer AI like Meituan Xiaomei.
More from Companies & People
- China’s AI scene is being led by startups, not internet giants — teortaxesTex · 2026-08-04
- Palantir’s FDE model is where AI becomes valuable inside enterprise systems — xiaohu · 2026-08-04
- Hugging Face CEO says open-weight models could help China pull ahead in AI — ben_burtenshaw · 2026-08-04
- Toby Walsh to launch new AGI risk book in Mosman on September 11 — TobyWalsh · 2026-08-04
- Sovereign AI critics are really just competing against frontier models — Afinetheorem · 2026-08-04
- AI clone ran a sales front line and surfaced about $3M in opportunities — nikola_mr64990 · 2026-08-04