AI BriefingLFM2.5-24B-A2B: Edge-optimized MoE model activating only 2B of 24B parameters
LiquidAI/LFM2.5-24B-A2B
About the project
It adopts a Mixture of Experts architecture with a total of 24 billion parameters, activating only 2 billion during inference. This enables efficient execution on edge devices or in resource-constrained hardware environments while maintaining the performance of large-scale models.
It supports 16 languages, including English, Korean, Japanese, and Chinese, and is specialized for conversational text generation. It includes built-in chat templates that support tool calling and reasoning process handling, allowing flexible adaptation to complex agent workflows.
It is compatible with major inference engines such as vLLM and SGLang, and provides an OpenAI-compatible API. It is suitable for developers looking to build high-quality LLM services in environments with limited computing resources, such as mobile or embedded systems.
LiquidAI/LFM2.5-24B-A2B
The original page has no description.
text-generation
This introduction was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report errors, attribution issues, or removal requests via Contact.