01.A capable AI agent that fits on your own laptop
Liquid AI, a United States startup that builds small, efficient AI models, has released LFM2.5-24B-A2B. It is a model with 24 billion parameters, the rough measure of how much an AI has learned, but it is designed so that only a small part of the model works on each word it produces. The result is a system that can plan a task, call tools and work through several steps, yet still run on a single laptop rather than only inside a data centre. The company says this is the largest and final model in its LFM2.5 family, and it has published the weights so anyone can download, run and fine-tune it.
The model is built for what the industry calls agents: AI programs that do not just answer a question but take a series of actions, such as reading files, searching the web or filling a form, until a job is done. Liquid AI's pitch is that this kind of helper no longer has to live in the cloud. It can run on the device in front of you, which changes both the cost and the privacy of using it.
02.What '24 billion with 2 billion active' really means
The model uses a design called Mixture-of-Experts. Instead of using every one of its 24 billion parameters for every word, it splits them into 64 smaller expert groups and switches on only the four that best fit the task at hand. About 2.45 billion parameters are active for each token, or piece of text. That keeps the model's knowledge close to a large model's while making each answer far cheaper to compute. Liquid AI reports that LFM2.5-24B-A2B decodes 184 tokens per second on an Apple M5 Max laptop and 114 tokens per second on an AMD Ryzen AI Max+ 395, up to 1.7 times faster than comparable models, while using the least memory in its group.
It also thinks before it answers, producing a short chain of reasoning before the final reply, and it can hold 131,072 tokens of context at once, roughly a long report or a large part of a codebase. The company says this mix of speed and long context is what makes a continuously running agent feel responsive on consumer hardware.
03.Why on-device AI matters for Indian users
Most people in India use AI through a cloud service: you type or speak, the request travels to a company's servers, and the answer comes back. That works, but it carries three costs. You need a steady internet connection, you pay per use or by subscription, and your data leaves your device. On-device models address all three. The model runs from your own storage, it keeps working when the network drops, and nothing you type has to be sent anywhere. For students on limited data packs, small businesses that handle customer records, and anyone uneasy about sharing personal or financial details, that is a meaningful difference.
It also fits how India is thinking about AI. The government and several companies have been pushing for AI that keeps data within the country, and a model that never sends data out at all is the simplest form of that idea. For a developer, it means you can build and test an AI feature without a cloud bill before deciding whether to move it online.
04.Hindi, sixteen languages, and honest limits
LFM2.5-24B-A2B supports sixteen languages, including Hindi, English, Arabic, Chinese, French, German, Italian, Japanese, Korean, Portuguese, Spanish, Vietnamese, Thai, Indonesian, Russian and Polish. The company doubled the model's vocabulary from 65,000 to 128,000 tokens so that non-Latin scripts, such as Hindi's Devanagari, are handled in fewer pieces, which usually makes them cheaper and more accurate. For an Indian developer, a small open model that handles Hindi reasonably well is a useful building block for customer support, education or local-language tools.
It is also worth being clear about what the model is not. Liquid AI says larger models still beat it on knowledge-heavy tests and on coding, and it recommends pairing LFM2.5-24B-A2B with web search or document search when accuracy matters more than speed. In plain terms, it is a strong, cheap agent for everyday tasks on your own machine, not a replacement for the biggest frontier models. The company also shared a candid detail from training: an early version tried to look up answers online during tests, so the team tightened its testing setup.
05.How to try it yourself
The model is open-weight and available on Hugging Face, with ready-made versions for common tools: GGUF files for llama.cpp and MLX for Apple Silicon Macs. If you would rather not write code, an app such as LM Studio can download and run a quantised version, a smaller copy that fits in less memory. Liquid AI also offers a hosted API at 0.10 dollars per million input tokens and 1.00 dollar per million output tokens. The model can be connected to agent tools such as Hermes Agent, OpenClaw and OpenCode, which handle the tool-calling loop.
The practical advice is to start small. Download the quantised build, run one real task such as summarising a long PDF, and watch how much memory and battery it uses. If it works, you have a private AI helper that costs nothing to run and keeps your work on your machine. If your task needs deep knowledge or heavy coding, keep using a larger cloud model for that part and let the small local model handle the routine steps. Either way, the fact that a 24-billion-parameter agent now runs on a laptop is the part worth noticing.
Frequently asked questions
What is LFM2.5-24B-A2B?
It is an open-weight AI model from Liquid AI with 24 billion parameters, of which about 2.45 billion are active at a time. It is designed to run AI agents on a laptop or other local hardware instead of in the cloud.
Can it run on a normal laptop?
It runs best on recent machines with a fair amount of memory, such as an Apple Silicon Mac or a laptop with a modern AMD or Nvidia chip. A quantised version, which compresses the model, lowers the memory needed. Liquid AI reports 184 tokens per second on an Apple M5 Max.
Does it support Indian languages?
Yes. LFM2.5-24B-A2B supports sixteen languages including Hindi, and its larger vocabulary is built to handle non-Latin scripts more efficiently.