If you've been paying monthly for a cloud AI assistant, or skipping it because your data is too sensitive to send anywhere, this week's release changes the math.
What's happening
Liquid AI shipped LFM2.5-2.6B this week: a 2.6-billion-parameter model built from the ground up for agentic tasks. It plans, calls tools (web search, file access, code execution), and handles multi-step workflows, all running entirely on local hardware. No cloud. No GPU required. It fits inside 2.5 GB of RAM and holds 30 tokens per second on a modern smartphone.
Liquid AI's Hugging Face release post covers the benchmarks in detail. Despite being under a third the size of 9-billion-parameter models, LFM2.5-2.6B outperforms them on tool-use tasks. It also carries a 128,000-token context window, enough to hold a full long report in memory at once.
The license is the practical highlight. The LFM Open License v1.0 allows any organization with less than $10 million in annual revenue to use this commercially, for free. Qualified nonprofits are explicitly exempt from even that threshold for research and non-commercial work. That is a license written for the orgs that most need help with AI costs.
Try this this week
- Install Ollama on your Mac or Windows laptop at ollama.com. One-click download, no account needed.
- Pull the model: open a terminal and run
ollama run hf.co/LiquidAI/LFM2.5-2.6B-GGUF. The Q4_K_M quantization is about 1.7 GB and runs well on any 8 GB laptop. - Try a real task. Draft a follow-up email, summarize a long supplier contract, or walk through a multi-step intake form. Anything you would normally paste into a cloud AI tool.
- Use it where cloud AI makes you nervous. Client notes, donor records, patient intake files. Nothing leaves your machine.
- Test your oldest laptop too. If it runs there, the whole team can use it without hardware upgrades.
The bigger picture
Cloud AI's pitch has always been "no setup." The trade-off is your data travels somewhere, the bill grows with usage, and you are one vendor change from starting over. On-device models flip all three. LFM2.5-2.6B is not the only option here (Ollama already has dozens of models to try), but it is one of the first where "agentic," "runs on a phone," and "free for small orgs" are all true at once. The smallest working thing sometimes fits in your pocket.