LFM2 24B A2B is a large-scale language model developed by Liquid AI, serving as the flagship model of the second-generation Liquid Foundation Models (LFM2) family. It employs a sparse Mixture-of-Experts (MoE) architecture designed to balance significant parameter capacity with efficient active compute. With 24 billion total parameters and approximately 2.3 billion active parameters per token, the model is optimized for high-throughput performance across both cloud and edge environments, including high-end consumer hardware.
Architecture and Design
The model utilizes a unique hybrid architecture that pairs efficient gated short convolution blocks with a small number of grouped query attention (GQA) blocks. Developed through hardware-in-the-loop architecture search, this design comprises 40 layers, specifically 30 double-gated LIV convolution blocks and 10 attention blocks. The MoE configuration features 64 experts per block with top-4 routing, a scaling strategy that allows the model to maintain a low inference latency and energy footprint while providing the capabilities of a much larger dense model.
Key Capabilities
LFM2 24B A2B is a general-purpose instruct model that supports native function calling, structured outputs, and web search integration. It is multilingual, providing support for nine languages: English, Arabic, Chinese, French, German, Japanese, Korean, Spanish, and Portuguese. Its architectural efficiency makes it particularly suitable for high-volume multi-agent workflows and high-throughput retrieval-augmented generation (RAG) pipelines where massive concurrency is required.
Prompting and Tool Use
For tool-based tasks, the model supports native function calling using specific special tokens. It is designed to generate Pythonic function calls encapsulated between <|tool_call_start|> and <|tool_call_end|> tokens by default. Liquid AI recommends providing tool definitions as a JSON object within the system prompt to guide structured output. Users can also request JSON-formatted function calls via specific system instructions to override the default Pythonic style.