LongCat
by Meituan · Model · Language & code models
unproven
Huge open model with only Meituan's benchmarks; the one hands-on test was unimpressive
MoE with 1.6T total and about 48B active parameters; takes text and images, built for long agentic coding
AccessOpen weightsSize1.6T parametersLicenseMIT (Fahd Mirza saw it on Hugging Face; no conflicting source)
Why we say so
Meituan's own table has it slightly ahead of Claude Opus 5 on BrowseComp and clearly behind on Terminal Bench, with the unstarred scores measured in-house. Fahd Mirza's tests were a weak web-page simulation and literal, not culturally grounded, translations. He also found access scarce, with free use limited and few API providers.
Evidence
- Nov 151.6T MoE with 48B active, native vision, coding and office tasks; via LongCat Chat and API, GitHub and Hugging Face marked coming soon
- Nov 15Given a doner kebab image and asked for a web page simulation, one run gave a poor, unconvincing result unlike the curated demos
- Nov 15Clearly behind Claude Opus 5 on Terminal Bench and Program Bench; unstarred scores measured in-house, only indicative