AI SYSTEMS

The model moved into 2.5 gigabytes.

Liquid AI put a 2.6 billion parameter model on Hugging Face on August 4 that runs in under 2.5 GB of memory. It makes no pictures at all, which is exactly why it is worth your attention this week.

THE SHORT VERSION

A capable text model now fits in about the space of a phone photo library, and runs without a network connection.

•  Liquid AI published LFM2.5-2.6B on Hugging Face on August 4. It has 2.6 billion parameters and runs in under 2.5 GB of memory.

•  Reported speed is 220 tokens a second on an Apple M5 Max and 113 on an AMD Ryzen processor. Those are the company's own numbers, measured on the company's own hardware.

•  It runs from day one in llama.cpp, MLX, vLLM, SGLang and ONNX, so most people already have something that will load it.

•  It is a text model. It makes no pictures, no video and no music. Read the rest anyway, because most of a picture job is not the picture.

Most weeks a model release changes nothing about how the work is done, and the honest thing is to say so. This one is worth a slower read, not because it improves on what you already use, but because of where it can sit. Two and a half gigabytes is small enough to live on the machine already on the desk, next to the folder the job is in.

Nothing here replaces the model that renders the picture. The interesting part is the other half of the day, the half made entirely of words, which has been quietly leaving the building on every job for two years.

The release is a company post on Hugging Face, which makes it a vendor announcement rather than an independent result. Both the instruction-tuned version and the base model are downloadable now. The claim the company leads with is that it holds up against models around four times its size on following instructions and calling tools, and it reports 80.07 on a multi-turn instruction benchmark where it puts its comparisons between 55.67 and 77.35.

Read that the way you would read any number a company publishes about itself. It picked the benchmark, the comparisons and the hardware. What is checkable without trusting anyone is the size and the format support, and those are the parts that matter here anyway.

The licence is the one thing the post does not spell out. Before anything of a client's goes near it, read the model card, because a permissive download is not the same as a permissive licence and the difference shows up in the contract, not in the demo.

It does not render an image, a frame of video or a bar of music. Nothing about this release moves generation onto the laptop, and a small text model is not a quiet route to a local picture model.

Nor is it the most capable writer available to you. A model this size trades range for the ability to sit in a corner of memory. On long reasoning, on unusual subject knowledge, on anything where the answer has to be right rather than merely well formed, the large hosted models remain better and it is not close.

What it is good at is the repetitive, structured, high-volume word work that surrounds a job. That work is most of the day and almost none of the credit.

Take one commissioned picture and count the things made of words. Rewriting the client's paragraph into a brief. Producing twenty variations of one line to test which phrasing holds. Naming and captioning the outputs. Reading the revision note and turning it into the one line that changed. Writing the alt text. None of that is the picture, and all of it currently travels to somebody else's server.

Moving that half onto the machine changes three things at once. The client's unreleased brief stops being uploaded, which is the part that actually matters when the work is under embargo. The cost of a variation drops to nothing, so trying twenty stops feeling wasteful. And it keeps working on a train, in a hotel, and on the day an interface changes underneath you.

The honest test is whether the task would survive a slightly worse writer. Sorting, restating and multiplying survive it. Judgment does not. Keep the judgment where the better model is, and let the volume come home.

Run this against your own last week before installing anything. If the answer comes back short, the release does not matter to you yet, and that is worth knowing before you spend an evening on it.

I will list the tasks I did on one recent job, in plain language. Sort every task into two groups. First, tasks that would still be acceptable if done by a noticeably weaker writer, because the task is sorting, restating, reformatting or producing many variations of something I will choose between myself. Second, tasks where a weaker writer would produce a worse result I could not catch by looking. For each task in the first group, say what the job actually needs as input and output, in one sentence each. Do not recommend any product. Do not tell me a task is suitable if my own description shows I relied on judgment I did not write down.

The last sentence is the one doing the work. Without it the answer will happily sort everything into the first group, which is the failure that makes people move real judgment onto a small model and quietly ship worse work for a month before noticing.

THE NEXT WORKING NOTE

Still finishing the local desk kit: the short list of word tasks I have actually moved onto the machine, what broke on each one, and the point where I moved two of them back.

Want it when it ships? Reply with send me the local desk kit and I will get it to you.

A QUESTION FOR YOU

What is still leaving your machine that probably should not be?

Reply with the one task you would most like to keep in the room, embargoed client copy or otherwise. If enough of you name the same one, that is the next issue.

If this was useful, forward it to someone paying by the token to rename files.