Back to Blog
    October 15, 20263 min read

    By Brian Hanson

    LFM-2.5: The Fast Way to Keep Your Data Off the Cloud

    TL;DR

    You can now run LFM2.5-24B-A2B locally on your computer using tools like LM Studio to perform automated tasks without sending data to the cloud or paying API fees.

    Key Takeaways

    • LFM-2.5 runs up to 1.7x faster than comparable models on-device.
    • The model scored 27.69 on AutomationBench, outperforming larger models at multi-step tasks.
    • Running locally eliminates data privacy risks and recurring costs per message.
    An illustration of a laptop on a wooden desk. The screen displays text about local processing and private results. A glowing 'LOCAL' icon sits next to the laptop, while a severed cable leads to a cloud symbol in the background. A brass padlock sits on the desk.

    Running AI agents usually means sending company data to a cloud provider and paying for every word the machine generates. If you want to automate your inbox or handle sensitive customer data, the wait times and lack of privacy can be a dealbreaker. Liquid AI just released LFM2.5-24B-A2B, a model built to solve this by running directly on your laptop or office workstation.

    This model uses a specific architecture that keeps only 2 billion (2B) parameters active at any time, even though it has 24 billion in total. Think of it like a massive library where the AI only needs to pull a few books off the shelf to answer a question. It hits speeds of 184 tokens per second on an Apple M5 Max (roughly 130 words every second). That is about 1.7x faster than other models in its class.

    Why Local Agents Matter for Your Business

    When you run an agent locally, you aren't waiting for data to travel to a server in another state. The response is nearly instant. More importantly, your data never leaves your physical machine. This is a safe way to handle payroll data, client contracts, or private strategy documents.

    LFM-2.5 is built for "agentic" tasks (a term for AI that is good at using other tools). It scored 27.69 on AutomationBench, which measures how well AI performs multi-step tasks. For comparison, the Qwen3.6-35B-A3B model scored 8.72 on the same test according to Liquid AI research.

    How to Get Started Today

    You don't need to be a developer to test this. You just need a tool that can host the model on your computer. Follow these steps to wire it up:

    • Download a Local Runner: Use a free application like LM Studio or Ollama. These programs act as the interface between you and the AI model.
    • Find the Model: Search for "LFM2.5-24B-A2B" within the app's search bar. Look for the version labeled "GGUF" to get the best performance on a standard Mac or PC.
    • Load and Chat: Once downloaded, select the model and start typing. It uses your computer's RAM (memory) to process your request.

    Because this model uses a "chain of thought" process, it shows you its reasoning before it gives a final answer. You can see if the AI is making a logical error before you trust the output. If the model lacks specific facts about your industry, you can pair it with your own local folders (a process called Retrieval) so it can search your documents for answers.

    If you want to see exactly how to stack these tools to automate your daily tasks, our free 3-day training walks through the setup step-by-step.

    FAQ

    What hardware do I need to run this?

    It runs on standard business hardware, reaching 184 tokens/s on Apple M5 Max chips and 114 tokens/s on AMD Ryzen AI Max+ 395 processors.