Liquid AI released LFM2.5-24B-A2B on October 13, targeting AI agents running on a local device. It published a base version and a version further trained for instructions, tool use and multistep tasks. Downloadable weights let developers run the model on hardware they control, rather than send every request to an external cloud model service during routine work.
The model has 24 billion parameters, but only some participate in each token’s computation. The A2B name and launch announcement round this to 2 billion active parameters; the model card specifies 2.45 billion. It uses a mixture-of-experts architecture, or MoE. The saving concerns active computation per step, not a promise that the entire model occupies as little memory as a model with only 2 billion parameters.
What changes with local inference
The context window is 131,072 tokens, and the model is designed to reason before its final answer. Liquid AI supports runtimes including llama.cpp, MLX and vLLM. In practice, quantization, context length and concurrent requests affect memory use. Developers therefore need a model file and runtime suited to their machine, rather than relying on the active-parameter count in the product name.
The company reports 184 tokens per second on Apple M5 Max hardware and advantages over larger models on several tool-use benchmarks. These are Liquid AI’s results under its test conditions, not a speed guarantee for another computer. Its own tables also show no universal victory: other models lead some knowledge and coding tests. The choice should match the task, beyond text-generation speed.
Privacy, licensing and language before deployment
Local inference can keep requests and model computation on the computer, but an agent is more than a model. Search tools, email services or business-system connections may still send information online. Before granting documents or permissions, specify allowed tools, where they send data and which actions need approval. Installing locally does not remove those checks.
Open weights do not mean unrestricted commercial licensing, either. LFM Open License v1.0 includes a commercial-use limitation tied to a 10-million-dollar annual-revenue threshold. Deploying organizations need to examine the terms that apply to them. The official language list includes English, Arabic, French and German, among others; Hebrew is absent, so this release does not establish reliable Hebrew support.
A practical starting point is one defined task with nonsensitive files and limited permissions, such as sorting documents or drafting from local sources. Test accuracy, task completion and resource consumption, not merely whether the model loads. The potential benefit is greater control over where computation happens. A reliable service still requires hardware matching, language testing and oversight of agent actions.