How much memory a local AI model needs

Before you install a local model, two numbers tell you whether it will run on the computer you already own: the size of the model file and how much memory (RAM) the computer has to hold it.

Base Drive is planned with two models, a 0.9 GB file for computers with 4 GB of RAM and a 3 GB file for computers with 8 GB.

What a 4-bit GGUF file is

A language model is a very large list of numbers called weights, or parameters. Labs often release those weights at 16 bits each, which works out to about 2 GB per billion parameters.

Quantization rounds each weight to fewer bits. A 4-bit, or Q4-class, file keeps about 4 to 5 bits per weight. Hugging Face's GGUF documentation (opens in a new tab) lists its Q4_K type at 4.5 bits per weight. That shrinks the file to roughly 30 percent of its 16-bit size, a little under a third: LFM2.5 1.2B (opens in a new tab) goes from 2.34 GB to 0.73 GB. Rounding gives up a little precision, and in return the model fits on a laptop.

GGUF is the single-file format that carries those weights plus the details needed to run them. The developer of llama.cpp created it, and llama.cpp, LM Studio and Ollama can all open it. Names like Q4_0 and Q4_K_M mark 4-bit variants. Every model size below is for a 4-bit file.

The two models on Base Drive

Base Drive is planned at $19.99 and cannot be ordered yet. As of October 2026 the site lists two GGUF models for it, a compact one and "the default everyday model, on every drive."

Liquid AI's model card (opens in a new tab) says the LFM2.5 family is designed for on-device deployment. Gemma 4 E2B is one of Google's small open-weight Gemma 4 models, described on its model card (opens in a new tab). The input and context rows come from the model cards.

The RAM row is the site's minimum for the whole computer. Liquid says the model itself "Runs under 1GB of memory."

Why parameter counts are only a rough guide

The "E" in Gemma 4 E2B stands for effective, and it shows how a headline number can mislead. Google's model card explains that each layer carries its own embedding table (per-layer embeddings), which takes space to store and is used only for quick lookups.

Divide the 3 GB file by the 2.3B effective parameters and you get about 1.3 GB per billion, roughly double the usual 4-bit rate. Divide it by the 5.1B total and you get about 0.6 GB per billion, in line with other 4-bit files.

The file stores every parameter, so when you check whether a model fits, trust the file size.

A memory ladder from the configurator

Our build-your-drive preview (opens in a new tab) sizes eight models. Two carry the badge "On the drive". The other six are marked Roadmap. The page is an early preview and its prices are placeholders, so treat its numbers as planning estimates. Here are the two drive models next to the configurator's three largest entries:

Any build without one of the three biggest models gets the label "CPU is fine". Add Qwen 3.5 9B or Gemma 4 12B and the preview asks for "a recent laptop with 16GB RAM. A GPU or Apple Silicon makes it smooth."

The preview's sizes are estimates built on roughly 0.65 GB per billion parameters, checked against published Q4_K_M files. Current Hugging Face files agree: at Q4_K_M, Qwen3.5-9B (opens in a new tab) is 5.68 GB, Gemma 4 12B (opens in a new tab) is 7.12 GB and Qwen3.6-27B (opens in a new tab) is 16.82 GB, about 0.6 GB per billion. Three rules of thumb follow:

• File size per billion parameters. Plan on about 0.6 to 0.7 GB per billion at 4-bit, so a 20B model lands around 12 to 14 GB. Gemma's "E" models break the rule.
• RAM runs about double the file or more. The extra memory holds the operating system, the chat app and the conversation.
• The preview lists GPU memory at or above each file size. That lets the model sit in graphics memory. The 27B entry has nothing to spare: the real Q4_K_M file is 16.82 GB, so a 16 GB card needs a smaller 4-bit variant or help from system RAM.

How the drive's storage adds up

Every drive carries a fixed 3 GB for the app, the offline docs and room for the memory file, which is marked coming soon. A Qwen3Guard 0.6B safety model adds 0.5 GB. It is planned for every drive, and the preview always counts it toward the total, even though it is not bundled yet.

The configurator preview's default custom build is Gemma 4 E2B alone (the $19.99 Base Drive itself carries both models):

For that total the preview recommends a 32 GB drive and shows 25.5 GB free. Adding LFM2.5 1.2B brings the total to 7.4 GB, which still fits.

The installer copies the model onto your laptop, so the laptop needs free disk space for the model file and enough RAM to run it.

How the installer picks a model for your computer

As designed, the installer does this math for you. According to How it works (opens in a new tab), it "verifies every file, checks the computer, and selects a local model that fits."

The compact model covers smaller machines. A computer with 4 GB of RAM falls short of Gemma 4 E2B's 8 GB and meets LFM2.5 1.2B's 4 GB, so it can still run a private assistant offline. A machine with 8 GB or more can run the stronger everyday model.

Smaller models come with a trade in accuracy. What a million real questions say about small local models covers where they do well and how to check their answers.