Liquid AI Releases LFM2.5-24B-A2B for Local Agent Workloads
Liquid AI is positioning a smaller-parameter open model against larger Google and Alibaba models on agent tasks specifically, and the tooling support arriving alongside the weights means the comparison can be checked locally rather than taken on faith.
Reporting from 1 source: GIGAZINE.
Liquid AI released LFM2.5-24B-A2B, a 24 billion parameter mixture-of-experts model with 245 million active parameters, pretrained on 33 trillion tokens and further trained for agent workflows. The company reports it scored higher than Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it, and Nemotron-3.5-Lightning (30B-A3B) on agent benchmarks including AutomationBench and BFCLv4. Inference support already exists in llama.cpp, MLX, vLLM, SGLang, and ONNX, and the model is downloadable from Hugging Face as an open model.
The model runs the Q4_K_M quantization at 184 tokens per second on a MacBook Pro with M5 Max, using 15GB of memory, and at 114 tokens per second on a PC with Ryzen AI Max+ 395. Liquid AI also published a demo in which Hermes Agent launches on a MacBook Pro with M5 Max and LFM2.5-24B-A2B carries out the task 'create a briefing about AI models that are trending on Hugging Face.' The weights are open and available on Hugging Face.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.