Liquid AI opens a reasoning model for agents running on local devices
Liquid AI has released the weights of LFM2.5-24B-A2B, a model that writes reasoning before its final answer and supports tools used in multi-step tasks. It extends the context and vocabulary of its predecessor while keeping the same architecture. Developers can download base and post-trained versions and connect a local model server to agent software through an OpenAI-compatible interface.
Artificial Intelligence··Night
Downloadable weights bring reasoning to local agents
The model developer Liquid AI released LFM2.5-24B-A2B on October 13 for agents that run on local devices. Base and post-trained weights are available through Hugging Face, a platform for sharing models. The release provides downloadable parameters that developers can run themselves. The model generates reasoning text before its final answer, and its training includes tool calls and tasks spanning multiple conversation turns.[1], [2]
The new version retains the architecture of LFM2-24B-A2B. Its mixture-of-experts design selects part of the model for each token, the units used to process text: approximately 2 billion parameters are active out of 24 billion in total. The architecture combines short convolution layers with attention, a mechanism that connects relevant parts of an input. Liquid AI has expanded the context from 32,000 to 131,000 tokens.[1]
Longer context accompanies expanded training
Vocabulary has grown from 65,000 to 128,000 tokens, with improved handling of non-Latin writing systems as a stated aim. Pretraining has increased from roughly 17 trillion to 33 trillion tokens. Intermediate training emphasizes reasoning, mathematics, code, instruction following and multilingual text. It also includes longer contexts and tool interactions, extending the training beyond isolated questions and answers.[1]
Post-training combines supervised fine-tuning, specialized teacher models, distillation and reinforcement learning for agent tasks. Liquid AI assessed the model across 16 benchmarks covering tool use, knowledge, instruction following, mathematics and coding. Those comparisons are the company’s own results. Its description also identifies knowledge-heavy tasks on which larger models retain an advantage.[1]
Local servers connect to agent software
Local execution support includes llama.cpp, MLX, vLLM, SGLang and ONNX. Developers can expose a local server through an OpenAI-compatible interface and connect it to agent software such as Hermes Agent, OpenClaw, Pi or OpenCode. This setup lets the model run without sending model requests to a remote service. Data sent by separately connected external tools depends on how those tools are configured.[1]
