Liquid AI发布了LFM2.5-24B-A2B模型，这是一个专为快速、私密的边缘设备智能体设计的推理模型1。该模型拥有24B参数规模，但仅有2B参数在推理时处于激活状态1，大幅降低了计算资源需求。模型的开源权重已发布至Hugging Face1。
在性能表现上，LFM2.5-24B-A2B在多个基准测试中与参数量更大的模型相竞争。在LiveClaw基准上，该模型得分为48.09，相比之下Qwen3.6-35B-A3B为38.191；在AutomationBench上，该模型得分27.69，而后者仅为8.721。运行速度方面同样表现出色，在Apple M5 Max上达到184 tokens/s的速度1，在单个NVIDIA B300 GPU上可达21K output tokens/s1，相对于竞品实现了1.7倍的速度提升1。
该模型在架构设计上进行了多项升级。上下文窗口从32,768扩展至131,072 tokens1，词汇表规模从65K扩展至128K tokens1。预训练数据量增加至33T tokens（相比之前的17T tokens）1，并通过中期训练的两个阶段进行微调，其中第一阶段使用2T tokens，第二阶段使用300B tokens1。该模型支持llama.cpp、MLX、vLLM、SGLang、ONNX等多种推理框架1，定价为输入令牌$0.10/1M（缓存$0.05）、输出令牌$1.00/1M1。
Liquid AI has released LFM2.5-24B-A2B, a 24-billion-parameter inference model designed for rapid and private edge device deployment 1. Despite its substantial parameter count, the model activates only 2 billion parameters during operation, enabling efficient inference on consumer hardware while maintaining competitive performance 1.
The model demonstrates strong benchmark results relative to larger alternatives, scoring 48.09 on the LiveClaw benchmark compared to 38.19 for Qwen3.6-35B-A3B, and achieving 27.69 on AutomationBench versus 8.72 for the competing model 1. Performance-wise, LFM2.5-24B-A2B reaches 184 tokens per second on Apple M5 Max processors and scales to 21,000 output tokens per second on a single NVIDIA B300 GPU, representing a 1.7x speed improvement over comparable alternatives 1. The model's context window has been expanded to 131,072 tokens from the previous 32,768, while its vocabulary was increased to 128,000 tokens 1. Pretraining extended to 33 trillion tokens, with intermediate training across two phases totaling 2.3 trillion tokens 1.
The open-source weights are available on Hugging Face and compatible with multiple inference frameworks including llama.cpp, MLX, vLLM, SGLang, and ONNX 1. Pricing for the model is set at $0.10 per million input tokens (with cached tokens at $0.05) and $1.00 per million output tokens 1.
评论
还没有评论，欢迎留下第一条。