Is 8 GB about VRAM or system RAM?

The catalog's numbers are RAM recommendations: memory gates whether a model runs at all, while a GPU mainly changes speed. Unified memory on Apple Silicon counts as RAM.

What do I give up at this size?

Depth on hard reasoning, long context, and reliability on multi-step tasks. The prefer-local toggle covers the gap: what fits runs on your machine, and the jobs that outgrow it fall back to the cloud catalog in the same chat.

How do I know a model fits before downloading?

Pair your computer in idapt and the hardware-fit filter marks what your actual RAM and GPU can serve, per quantization, from the same data this page is built on.

All roundups

Best of

Real local AI on 8 GB.

Every pick's quantization recommends at most 8 GB in the catalog's hardware data: what fits, what it trades, how to run it.

Last reviewed July 16, 2026

Try idapt free Browse all models

8 GB is the most common laptop spec and the tightest local tier: the line between a model that answers and one that makes the machine swap. Every pick below has an Ollama variant whose recommended RAM is at most 8 GB in the catalog's hardware data, so this page is scoped to what actually runs. If your machine has 16 GB or more, the general local roundup covers the larger classes.

How we pickedPicks are filtered by the catalog's per-variant local-inference data: a model qualifies only if a variant's recommended RAM is 8 GB or less (the same numbers the 'runs on my machine' filter enforces). At this tier that means 3B to 9B models at 4-bit quantization, so each entry states the trade its size buys.

The picks

1Qwen3 8BThe current-generation default.

$0.05 in / $0.4 out per M tokens · 41K context

The newest generation at this size: the q4 build downloads about 5 GB and recommends 8 GB of RAM, with the strongest instruction-following of the picks here.

Best for: The default chat model on an 8 GB machine.

2Llama 3.1 8BThe ecosystem standard.

$0.02 in / $0.03 out per M tokens · 131K context

The small model the local ecosystem standardized on: tutorials, fine-tunes, and tooling assume it, and its q4 build recommends exactly 8 GB.

Best for: Compatibility with the wider local tooling world.

3Qwen2.5 Coder 7BThe coding pick that fits.

Code-tuned at a size that leaves room for your editor and browser; quick local completions and small fixes without a cloud round-trip.

Best for: Local code help on 8 GB machines.

4Llama 3.2 3BThe headroom pick.

$0.051 in / $0.34 out per M tokens · 80K context

Recommends just 4 GB (6 GB at q8), so it runs while the browser, editor, and a meeting stay open; the practical choice when 8 GB is the whole machine, not the model's budget.

Best for: 8 GB machines that keep multitasking.

Frequently asked

Run them in one workspace

Local AI

Run open models on your own hardware with Ollama: private, free, no rate limits.

Computers

Daemon-connected machines with real filesystems and agent-friendly controls.

Keep comparing

Serious models, your hardware.Open weights at the frontier.The models that cost nothing to run.idapt vs Jan idapt vs LM Studio

Judge them on your own prompt

Every pick on this page runs in one idapt chat. Start free with a daily allowance, no credit card.

Try idapt free