88Z AI, 10 November — Meituan's LongCat team has open-sourced the LongCat-2.5 model under the MIT license, allowing commercial use. With 1.6 trillion total parameters and approximately 48 billion activated, it supports million-token contexts and long-horizon agent tasks. Optimized for domestic computing power, it significantly improves single-card training throughput.
At a glance
- LongCat-2.5 is open-sourced under MIT license, supporting million-token contexts and long-horizon agent tasks
- Optimized for domestic computing power, achieving 86% higher single-card SFT throughput than DDP
The update
The LongCat team under Meituan has officially open-sourced the LongCat-2.5 model. Released under the MIT license, it fully supports commercial inference and further training. The model features 1.6 trillion total parameters with approximately 48 billion activated, capable of handling million-token contexts and designed specifically for long-horizon agent tasks. Technically, it employs W4A4 quantized inference and is deeply optimized for domestic computing hardware. It enables long-sequence training even when single-card memory is under 80GB. Tests show that its supervised fine-tuning (SFT) single-card throughput is 86% higher than distributed data parallel (DDP) methods, effectively lowering deployment and training barriers.
Sources & editorial note
This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.
- 龙猫LongCatSource published
Source dates above refer to the original announcements; the date under the headline is this article’s publication on 88Z AI.