Local AI

Run language models and image generation on your own hardware. The platform AI watches and meters; the local AI runs in your house. For speed, privacy, and zero-marginal-cost workloads — not frontier reasoning.

16 tools · last updated 2026-07-26

AUTOMATIC1111

free · open source

AUTOMATIC1111's Stable Diffusion web UI is the long-standing default interface for generating images locally with Stable Diffusion models. It runs in your browser against your own GPU, with a huge extension ecosystem and simpler controls than node-based tools like ComfyUI. Your prompts and images never leave your machine. Development has slowed, but it remains the easiest local image-generation on-ramp.

Rung 5 · Stack/ Compute Trustless

ComfyUI

free · open source

ComfyUI is an open-source, node-based editor for running image-generation models like Stable Diffusion and Flux on your own GPU. You wire models, prompts, and processing steps into reusable workflows, which makes it the professional's tool for local image work. Prompts and outputs never leave your machine. The node interface has a real learning curve compared to simpler front ends.

Rung 5 · Stack/ Compute Trustless

Family AI Chat

freemium

Family AI Chat is a closed-source mobile app offering push-to-talk voice conversations with an AI companion, where an adult configures allowed topics, voices, and languages. Speech is transcribed on the device and conversation text is kept locally, but the text itself goes to an unnamed third-party AI service for processing, so you cannot evaluate who actually handles it.

Rung 0 · Default/ Compute Trusted third party

GPT4All

free · open source

GPT4All is a free, open-source desktop application from Nomic that runs open-weight chat models on ordinary consumer hardware with essentially no setup. Install it, pick a model, and everything — prompts, documents, answers — stays on your machine, including a local-document chat feature. It will not match frontier cloud models for hard reasoning, but private everyday tasks cost nothing per use.

Rung 4 · Operate/ Compute Trustless

HomeLearnAI

free · open source

HomeLearnAI is an MIT-licensed Laravel application for homeschool management, with spaced-repetition review scheduling, multi-child support, calendar integration, and separate adult and learner interfaces. You can run the hosted version free or deploy it on your own server. The published documentation never states where its AI inference runs, so self-hosting the app does not confirm you are self-hosting the model behind it.

Rung 2 · Custody/ Compute Hybrid

Hugging Face

freemium · open source

Hugging Face is the central registry for open-weight AI models — the place where models you can download and run on your own hardware are actually published. Every local tool on this list pulls from it directly or indirectly. It is a venture-backed platform, not sovereign infrastructure itself, but once the weights are on your disk, the platform's fate no longer affects you.

Rung 5 · Stack/ Compute Hybrid

Jan

free · open source

Jan is an open-source desktop app that gives you a ChatGPT-style chat interface running entirely on your own computer, with no account and no telemetry. Download a model, start chatting; conversations never leave the machine. It suits people who want the familiar chat experience without a platform watching, and accepts the usual local-model tradeoff: less reasoning power than frontier cloud models.

Rung 4 · Operate/ Compute Trustless

llama.cpp

free · open source

llama.cpp is the open-source C++ inference engine underneath much of the local-AI ecosystem, including Ollama and LM Studio. It runs open-weight models efficiently on ordinary CPUs and GPUs and gives you direct control over quantization, context size, and performance. It is a power tool: most people should start with the friendlier layers above it and drop down only when they need that control.

Rung 5 · Stack/ Compute Trustless

llamafile

free · open source

llamafile, a Mozilla project, packs a language model and its runtime into a single executable file that runs on macOS, Windows, Linux, and the BSDs with no installation at all. Copy the file to any machine — even one with no internet — and you have working private AI. It is sovereignty by way of portability, better as a demo and utility than a daily driver.

Rung 4 · Operate/ Compute Trustless

LM Studio

free

LM Studio is a polished desktop app for running open-weight language models locally, with a built-in model browser and one-click downloads from Hugging Face. It is the easiest path for non-developers: no terminal required, and your prompts stay on your machine. The app itself is proprietary, which is a real if modest tradeoff against fully open alternatives like Jan or Ollama.

Rung 4 · Operate/ Compute Hybrid

LocalAI

free · open source

LocalAI is an open-source, drop-in replacement for the OpenAI API that runs entirely on your own hardware. One self-hosted endpoint serves chat models, embeddings, speech, and image generation, so software written for cloud AI can point at your server instead with no code changes. It is aimed at people running a home server or homelab rather than a single desktop app.

Rung 5 · Stack/ Compute Trustless

MaxHome AI

freemium

MaxHome AI is a closed-source household assistant app for iOS and Android with per-person profiles, voice commands, calendars, reminders, grocery lists, and budget tracking. It states conversations are never sold to advertisers or used to train models. As shipped it is a cloud service; a fully local install exists only through a paid in-home consultation, so the default configuration sends your household data off-device.

Rung 0 · Default/ Compute Trusted third party

MLX

free · open source

MLX is Apple's open-source machine-learning framework built specifically for M-series Macs, using unified memory to run open-weight models that would not fit in a typical graphics card's memory. For a Mac owner, it is the native fast path to serious local inference on hardware you already own. It is developer-oriented; non-coders will reach it indirectly through apps like LM Studio.

Rung 5 · Stack/ Compute Trustless

Ollama

free · open source

Ollama is the simplest way to run open-weight language models on your own computer: one command to install, one to run a model. It exposes an OpenAI-compatible API on localhost, so tools built for cloud AI can work against your own hardware instead. Prompts never leave your machine. Local models trail the frontier in reasoning, so match them to the right jobs.

Rung 4 · Operate/ Compute Trustless

Open WebUI

free · open source

Open WebUI is a self-hosted, ChatGPT-style web interface that sits on top of Ollama or any OpenAI-compatible endpoint. It adds the product layer local AI usually lacks — chat history, multiple users, document chat — while logging nothing to anyone else's servers. You run it yourself, typically in Docker, so it suits people one step past the beginner stage who want local AI to feel finished.

Rung 5 · Stack/ Compute Trustless

vLLM

free · open source

vLLM is an open-source serving engine for running language models at high throughput on real GPU hardware, using techniques like PagedAttention and continuous batching to squeeze maximum performance from each card. It is the production tier of self-hosting: the right tool when you are serving a team, an app, or paying customers, and overkill for one person chatting on a laptop.

Rung 5 · Stack/ Compute Trustless

Frequently asked questions

How much GPU do I need to run a local LLM?
It depends on model size. 8GB of VRAM runs small 7–8B models; 16GB handles everyday chat and RAG (feeding the model your own documents) with 13–14B models; 24GB is the sweet spot for the strong 27–40B class. Anything around 70B needs 24–48GB even after quantization (compressing the model to fit). More VRAM buys bigger models — not necessarily better answers for your actual task.
Can I run an LLM locally without a GPU?
Yes. Small quantized models (7B and under) run on a plain CPU, and Apple Silicon Macs run mid-size models well because the M-series chips share system memory with the graphics cores. Expect a real speed tradeoff: CPU-only generation often crawls at a few tokens per second — fine for short answers, painful for long documents. For daily use, a unified-memory Mac or a modest GPU makes the difference.
Why are local LLMs so slow?
Usually memory bandwidth, not processing power. Generating each token means reading the entire model from memory, so speed is capped by how fast your hardware moves data — which is why a GPU with fast VRAM beats a stronger CPU, and why Apple Silicon punches above its specs. Fixes: use a smaller or more quantized model, keep it fully in VRAM, and trim your context. Some slowness is the price of ownership.
What is the best local LLM for coding?
There's no permanent answer — the leaderboard turns over every few months. As of mid-2026, Qwen's coder models and DeepSeek's coding releases lead most local benchmarks, with 30B-class versions running well on a 24GB GPU. Whatever you pick, run it through Ollama or LM Studio and test it on your own codebase; benchmark rank rarely predicts how a model handles your stack. Expect it to trail frontier cloud models on hard problems.
Is a local LLM private?
Mostly, yes — and that's the point. Once the model weights are on your disk, prompts and outputs never leave your machine, and no provider logs your questions. Two caveats: downloading a model reveals what you fetched to the registry that hosts it, and some runner apps ship with update checks or telemetry you should disable. Run it offline after setup and the privacy claim holds completely.