AI BriefingMeituan unveils 1.6T-parameter MoE model
Key point
Meituan has released LongCat-2.0, an open-source MoE model with 1.6T parameters trained on AI ASICs.
Details
LongCat-2.0, developed by Meituan, is a MoE (Mixture-of-Experts) model with a total of 1.6T (1.6 trillion) parameters, using approximately 48B (48 billion) active parameters during inference.
Key features of this model include:
- Support for a 1M (1 million) token long context window
- Application of LongCat Sparse Attention technology
- Post-training optimized for coding and agentic workflows
- Integration with Claude Code, OpenClaw, Hermes, and others
Notably, the technical highlight of this model is that, moving away from the conventional Nvidia GPU-centric training approach, it was trained on over 35T (35 trillion) tokens of data using AI ASIC Superpods, completing stable training with no loss spikes throughout the process. The model is provided as open weights under the MIT License.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
Jevper Released: A Jev Classification Wrapper for OpenAI-Compatible Models with Automatic Logprobs Detection
Rootless Linux Sandbox 'Drop' Released for Isolating AI Agents and Third-Party Programs