PANews June 30 news, according to Jin10 reports, today, Meituan officially released its new generation trillion-parameter large model LongCat-2.0, and will open-source it. LongCat-2.0’s pre-training data scale exceeds 30T tokens, covering Chinese, English, multilingual, code, and other data types; facing hardware failures, communication anomalies, memory pressure, and numerical fluctuations in training across tens of thousands of cards, the LongCat team tackled the challenges of domestic computing power training from three aspects: stability, correctness, and efficiency. In terms of stability, through HCCL exception handling, elastic card scaling, and automatic fault recovery, the average daily failure rate per month was reduced by over 70%; in terms of correctness, by self-developing deterministic operators, Bitwise consistency verification, and parameter detection, training result reliability is guaranteed, while also improving computational precision of key modules and optimizing Reduce logic based on practice.
Meituan Releases Next-Generation Trillion-Parameter Large Model LongCat-2.0
If you have any feedback or questions about this content, please contact us at crypto.news@kcex.com
Disclaimer: The articles reposted on this website are sourced from public platforms and are for reference only. These articles do not represent the views or opinions of KCEX. All copyrights belong to the original authors. If you believe that any reposted article infringes upon the rights of a third party, please contact crypto.news@kcex.com for removal. KCEX makes no representations or warranties regarding the timeliness, accuracy, or completeness of reposted articles, and shall not be liable for any actions or decisions made based on such content. Reposted materials are for informational purposes only and do not constitute advice, endorsement, or basis for any commercial, financial, legal, and/or tax decisions.