AI News (2026/11/10): Meituan LongCat Team Officially Open Sources LongCat-2.5
Executive Summary:
On November 10, Meituan's LongCat team released the open-source version of LongCat-2.5, achieving a breakthrough in domestic computing power compatibility while maintaining the openness of the MIT license. This model constructs million-level context processing capabilities with a total of 160 billion parameters and 48 billion activated parameters, and reduces deployment barriers through W4A4 quantization technology.
News Details
Meituan's LongCat team continues to deepen its expertise in long-sequence modeling, launching a new version model after LongCat-2.0. This open-source version features architectural optimizations for agentic workflow scenarios and has achieved a complete training loop on domestic GPU clusters with single-card VRAM below 80GB.
[160 Billion Parameters]
LongCat-2.5 adopts a Mixture of Experts (MoE) architecture design, expanding total parameter scale to 160B level while maintaining model density. Its activated parameter count is controlled within 48B range, demonstrating superior computational efficiency in long-text understanding and generation tasks.
[Million-Level Context Window]
Through improved positional encoding mechanisms and attention calculation strategies, this model extends context length to 1 million token level. Practical tests show approximately 37% accuracy improvement in tasks requiring global semantic awareness such as document analysis and code review.
[W4A4 Quantization Deployment]
The W4A4 mixed-precision solution developed for domestic GPU characteristics reduces inference latency to one-third of traditional FP16 mode. The team simultaneously released supporting toolchains that enable automatic conversion from HuggingFace format to domestic computing platforms.
AI-ALL In-Depth Analysis
The open-sourcing of LongCat-2.5 marks a critical step forward for domestic large model technology stacks toward production-level applications. Its million-level context capability directly addresses information fragmentation issues in cross-document reasoning scenarios - developers can now complete semantic modeling at book-level without manual fragment stitching via RAG technology.
The MIT license's commercial application openness will accelerate enterprise adoption of this model. Particularly noteworthy is its foundational optimization for domestic computing platforms that may become industry benchmark cases. However, while single-card throughput improved by 86%, at least eight V100-level hardware units are still required to support complete training workflows.
This open-source release presents dual challenges to existing AI infrastructure: first validating domestic GPUs' feasibility boundaries for ultra-large-scale model training; second exposing room for improvement in current distributed training frameworks' resource utilization efficiency - balancing high throughput with hardware cost control will become one of the key research directions moving forward.
