Meituan Open-Sources Trillion-Param LongCat-2.0
美团技术团队 · wechat · 2026-07-09
Meituan's tech team announced the open-sourcing of LongCat-2.0. The model has a total of 1.6T parameters with an average activation of about 48B, and is available in multiple precision formats including BF16, FP8, and INT8. The team stated that the model has completed inference on a 50,000-card domestic compute cluster, focusing heavily on real-world Agentic Coding tasks, with model weights and inference code released simultaneously.
The post details LongCat-2.0's optimizations for domestic chip adaptation, KV-cache transmission, PD separated deployment, long-context processing, and sparse attention. It also covers post-training strategies for execution, reasoning, and interaction, emphasizing the goal of making a reproducible inference stack accessible to more domestic chips and existing compute resources.
Related event: Meituan Open-Sources LongCat-2.0, a Trillion-Parameter MoE Model(2 posts)→
More from Infra
- Encrypted GPT-2 speedrun challenges FHE inference: best is 3.031 s/token, goal sub-1s — tensorqt · 2026-10-27
- Open-source Latch compiles PyTorch into training megakernels, 2.75x faster on H100 — sguada · 2026-10-27
- NVIDIA claims HGX Rubin NVL8 delivers 10x tokens per megawatt vs B200 — thursdai_pod · 2026-10-27
- NVIDIA Claims HGX Rubin NVL8 Delivers 10x More Tokens per Megawatt Than B200 — thursdai_pod · 2026-10-27
- Together lets multiple LoRA post-training experiments share the same GPU capacity — togethercompute · 2026-10-27
- Commoditization dilemma: either human labor or GPUs become the commodity — _AustinCalvert_ · 2026-10-27