LongCat-2 report says 1.6T model was trained entirely on domestic chips
ThePrimeClock · reddit · 2026-07-23
A Reddit thread dissects the LongCat-2 technical report and concludes that Meituan is publicly demonstrating a frontier-scale model trained and served entirely on domestic Chinese silicon, without Nvidia GPUs.
The post cites the report’s own Chinese meta description, which says the 1.6-trillion-parameter model was trained completely on domestic chips. It also highlights several infrastructure details: 50,000+ ASICs for pretraining, tens of thousands grouped into 48-chip “superpods,” RoCE networking between pods, and a two-tier design that reportedly lifts throughput by about 30%.
The thread also notes the software compensations needed for less memory per chip than an H800: ZeRO-1 sharding, recomputation, activation offloading, custom deterministic operators, binary-tree reduction to limit floating-point error, bit-flip detection, and fault isolation for bad network links.
More from Companies & People
- ROS names Alejandro Hernández Cordero new project lead, eyes 2027 release — ___Mufasaa · 2026-10-27
- Anthropic S-1 reveals Dario can handpick a board majority forever via founder control — cocktailpeanut · 2026-10-27
- Mistral researcher: research freedom is overrated compared to company problems — konstmish · 2026-10-27
- ROS Meetup coming to Google's NYC office Oct 29 with speakers from Tangent, Cohesive Robotics, Intrinsic — kscottz · 2026-10-27
- All 8 Transformer paper authors left Google, and rival labs captured most of the frontier value — signulll · 2026-10-27
- Anthropic's S-1 pledges AI for curing diseases and serving 4.6B underserved patients — petergyang · 2026-10-27