What does my machine need?

Roughly: 8 GB of VRAM or unified memory runs small models, 16 GB runs the efficient mid-sizes, 24 GB+ runs the 27B class comfortably. Pair your computer in idapt and the hardware-fit filter answers precisely for your hardware.

What happens when a job is too big for the local model?

Routing handles it: 'prefer local' sends what fits to your machine and falls back to the cloud catalog for the rest, in the same conversation.

Is local really private?

Local inference runs on your machine through the idapt daemon; prompts and outputs stay on your hardware for those requests. The privacy page documents exactly what does and does not leave.

All roundups

Best of

Serious models, your hardware.

The local models worth running via Ollama, by hardware tier, with a cloud fallback when the job outgrows the machine.

Last reviewed July 16, 2026

Try idapt free Browse all models

Local models buy privacy (prompts never leave the machine), zero marginal cost, and offline capability; the price is capability per size. The picks below are the models worth their VRAM at each tier. Paired with idapt's local inference, they route like any catalog model: the hardware-fit filter shows what your machine can actually serve, and one toggle prefers local with cloud fallback.

How we pickedPicks are open-weight models with strong small-size benchmark standing, organized by the hardware they realistically need. The 'runs on my machine' fit comes from idapt's hardware-fit matching against your paired computer, not from this page's guesses.

The picks

1Gemma 4 31BThe workstation default.

$0.12 in / $0.35 out per M tokens · 262K context

The most capability currently packed into a size one strong GPU can serve; excellent instruction-following for its class.

Best for: 24 GB-class GPUs and Apple Silicon with 32 GB+.

2Mistral Small 3The efficient all-rounder.

$0.15 in / $0.6 out per M tokens · 262K context

Strong multilingual capability with efficient serving, comfortable on mid-range hardware.

Best for: 16 GB-class machines and multilingual work.

3Qwen 3.7 PlusThe capable middle of the open family.

$0.32 in / $1.28 out per M tokens · 1M context

The Qwen family's mid-size delivers a large share of the flagship's behavior at a size serious local rigs can serve.

Best for: Bigger local rigs that want near-flagship behavior.

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

Open weights at the frontier.Frontier answers at commodity prices.Real local AI on 8 GB.idapt vs Open WebUI idapt vs Msty

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