LongCat-Flash-Lite - Meituan's open-source large language model using LongCat.

What is LongCat-Flash-Lite?

LongCat-Flash-Lite is a new generation of high-efficiency large-scale language model launched by Meituan. The model employs an innovative hybrid expert (MoE) + N-gram embedding architecture, with a total of 68.5 billion parameters. Only about 2.9 to 4.5 billion parameters are activated per inference iteration, balancing powerful capabilities with extremely high efficiency. The model supports ultra-long contexts of up to 256K characters and performs exceptionally well in core evaluations such as agent tasks, code generation, and mathematical reasoning. In tool invocation and programming domains, its performance reaches the top level among models with similar activation scales. Through dedicated system optimization, the model's inference speed has been significantly improved.

Main functions of LongCat-Flash-Lite

• Text generationThe model supports multi-turn dialogue interaction and can generate fluent and coherent natural language responses.
• Tool callThe model has function call capabilities and can autonomously use external tools to complete complex tasks.
• Code generationProficient in programming tasks, able to write, understand and debug code in multiple programming languages.
• Long context processingThe model supports 256K ultra-long contexts, enabling it to handle tasks such as long document analysis.

The technical principles of LongCat-Flash-Lite

• MoE + N-gram Embedded ArchitectureLongCat-Flash-Lite employs a hybrid expert (MoE) architecture with a total of 68.5B parameters, activating only 2.9B to 4.5B parameters. Unlike traditional MoEs that use FFNs as experts, this model innovatively introduces N-gram embedding tables (NEs) to replace some experts, forming a hybrid architecture of "MoE + NE".
• N-gram embedding table mechanismN-gram embedding tables transform some computationally intensive operations into table lookup operations by pre-computing and storing embedding vectors composed of N-gram combinations. The core advantage of this mechanism lies in its more regular memory access patterns and lower latency for embedding lookups, significantly reducing data movement overhead compared to FFN forward propagation. Through systematic expansion experiments, the research team determined key factors such as the optimal integration timing, parameter budget allocation, hash collision mitigation strategies, hyperparameter configuration, and embedding initialization methods for the embedding tables.
• Inference efficiency optimization systemTo fully leverage the performance advantages of N-gram embedding tables, the team introduced a dedicated inference optimization system. The system comprises two core components: an N-gram Cache, which reduces the overhead of redundant embedding lookups through intelligent caching strategies; and synchronized kernels, custom-designed CUDA kernels that enable pipelined parallelism between embedding lookups and other computations.
• Long Context Extension TechnologyThe model employs YaRN (Yet another RoPE extension method) to extend the context length to 256K. YaRN improves Rotary Position Embedding (RoPE) by dynamically adjusting the rotation angle and temperature scaling factor of the position encoding. This allows the model to effectively handle longer input sequences without additional training, based on a shorter context used during training (typically 4K-8K), while maintaining sensitivity and stability to relative positional relationships.

LongCat-Flash-Lite project address

• HuggingFace model libraryhttps://huggingface.co/meituan-longcat/LongCat-Flash-Lite

Application scenarios of LongCat-Flash-Lite

• Intelligent Customer Service and Dialogue SystemThe model supports multi-turn dialogues and tool calls, and is suitable for intelligent customer service scenarios in industries such as aviation, retail, and telecommunications. It can autonomously query information, process orders, and resolve user issues.
• Code development and assisted programmingIt performs excellently on code benchmarks such as SWE-Bench and can be used for code generation, bug fixing, code review, and automated programming tasks to improve development efficiency.
• Long document processing and analysisThe model is applicable to scenarios that require processing large amounts of text, such as legal contract analysis, academic paper reading, batch processing of financial statements, and understanding long video scripts.
• Intelligent agents and automated workflowsThe model's powerful agentic capabilities can serve as the core engine to drive automated workflows and execute multi-step tasks, such as data collection, report generation, and system maintenance.
• Multilingual content processingThe model performs well across multiple languages, including Chinese and English, and is suitable for global business scenarios such as cross-language content generation, localization translation, and multilingual knowledge question answering.