---
title: "Local coding setups by machine"
description: "The machine classes with a passing local coding setup: tested hardware, pinned model, download size, and what runs today."
canonical: https://oloop.run/local
updated: 2026-10-05
---

# Local coding setups we measured

We measured one setup on each of 3 machine classes. Each page lists the hardware, the pinned model, and the run results.

## Measured machine setups

The machine classes with a passing qualification record, and the setup we tested on each.

| Machine | Tested setup | Model | Download | Page |
| --- | --- | --- | --- | --- |
| Apple Silicon 64–128 GB | apple-silicon-64-128@1.5.2 | mlx-dspark/qwen3-8b-8bit | [Read the measured setup](/local/apple-silicon-64-128) |
| Linux x64 NVIDIA L4 16–64 GB | linux-x64-l4@1.2.2 | llamacpp/qwen3-8b | [Read the measured setup](/local/nvidia-l4) |
| Linux x64 CPU 8–32 GB | linux-x64-cpu@1.2.2 | llamacpp/qwen3-1.7b | [Read the measured setup](/local/linux-cpu-8-32) |

A class with only a pending record is not listed. /setups carries every published setup, measured or not.

## What runs today

An explicit setup runs on the OS its catalog entry declares, after that entry's RAM, VRAM, and driver checks. The no-arg bootstrap is narrower.

What the engine supports today.

| System | What works | How |
| --- | --- | --- |
| macOS, Apple silicon | No-arg auto bootstrap | oloop host start picks a qualified setup for the machine class. |
| Linux x64 | Explicit setup | oloop host start --setup <id>@<version>. No-arg bootstrap stays macOS-only. |
| Windows, native | No oloop setup | The native local AI stack (Windows ML, Foundry Local, Phi Silica) is explained in the Windows guide. |
| Windows, WSL2 | The CLI, using the Linux path | Installs the POSIX CLI; it is not a native host. |

The full rule is on /docs#compatibility.

- [Setup compatibility](/docs#compatibility)
- [The Windows guide](/guides/windows-local-coding-agent)

## Not measured yet

Configurations with no measured record.

| Configuration | Status | Where to look |
| --- | --- | --- |
| 64 GB VRAM | Not measured | The catalog lists no setup at that VRAM size. |
| RTX 4090 | Not measured | No passing record for that card. |
| Mac with 32 GB | Not measured | The Apple page covers 64–128 GB; 32 GB is a different class. |
| Multi-GPU | Not measured | Every measured row uses one GPU. |

See /setups for the full catalog and its evidence.

- [All machine setups](/setups)
- [The memory guide](/guides/local-coding-model-by-memory)

## Install

Install the CLI, then start a host with the setup for your machine class.

```sh
curl -fsSL https://oloop.run/install.sh | sh
oloop host open
```

- [Install and requirements](/docs#install)

## Guides

These guides cover choices the catalog cannot measure yet.

- [Why own the agent runtime](/guides/why-own-the-runtime)
- [Local coding model by memory size](/guides/local-coding-model-by-memory)
- [Windows local AI: the stack and a coding agent](/guides/windows-local-coding-agent)
- [Self-hosted coding agent for a team](/guides/self-hosted-coding-agent)
- [Secure local coding agent](/guides/secure-local-coding-agent)

## Questions

### Why do only 3 machine classes have measured setups?

A page exists only when the catalog entry has a passing qualification record for that class. A pending or missing record stays a gap.

### What does a machine page prove?

One pinned setup passed one probe on the stated hardware and date. It is not a security review or a quality score.

Updated 2026-10-05 · Outerloop