Decentralized / Local LLM Hosting Overview

This page is an effort to collate and capture the state of various local & community LLM hosting projects, in an effort to help guide refugees from the large centralized players. The initial collection of projects was initially generated with an LLM (Claude Sonnet high thinking) so please consider adding any personal experiences, advocacy, warnings, etc.
If your nontechy neighbour asks about AI, what would you recommend?

Local LLM Tools Comparison

Project Category Mission / Positioning Platform(s) Engine/Backend Model Formats Interface API Compatibility License
Ollama Inference server Simple, single-command local model management — the de facto default backend many other tools build on macOS, Windows, Linux, Docker Own engine (llama.cpp-derived) GGUF, Modelfile CLI-first; minimal official GUI, large third-party UI ecosystem OpenAI-compatible + native API Open source (MIT)
LM Studio Desktop chat app Polished, beginner-friendly app for discovering, downloading, and running local models macOS, Windows, Linux llama.cpp, MLX (Apple Silicon) GGUF, MLX Full native chat GUI with model browser Local OpenAI-compatible server; MCP tool support Free, proprietary
Jan Desktop chat app Open-source, privacy-first “ChatGPT alternative” that runs entirely on your own machine macOS, Windows, Linux llama.cpp GGUF Native chat app with assistants/threads Local OpenAI-compatible server; optional cloud provider passthrough Open source (AGPLv3)
Osaurus Inference server / edge runtime Lightweight, native Apple Silicon–only LLM server, expanding into a full agentic “AI edge runtime” macOS (Apple Silicon only) MLX, Apple Foundation Models (macOS 26+) MLX Built-in chat UI; server-first OpenAI, Anthropic, and Ollama-compatible endpoints Open source (MIT)
Tiles Desktop chat app Private, collaborative AI assistant built on AT Protocol, emphasizing user-owned identity and encrypted peer-to-peer sync macOS (Apple Silicon), Linux (NVIDIA) — public alpha llama.cpp, embedded “Pi” agent harness GGUF Native app; browser WebUI planned Open source (MIT / Apache 2.0)
GPT4All Desktop chat app Fully offline, private AI that runs on ordinary consumer hardware, including CPU-only macOS, Windows, Linux llama.cpp-based, Vulkan (CPU) GGUF Native chat app with built-in document (LocalDocs) RAG Local API server Open source (MIT)
Msty Studio Desktop chat app Polished multi-model desktop client with split-view chat and local + cloud model switching macOS, Windows, Linux llama.cpp, Ollama GGUF Native chat GUI Local API server Freemium (paid “Aurum” tier)
text-generation-webui (oobabooga) Desktop chat app Maximally configurable local chat interface for sampler-level and format tuning macOS, Windows, Linux llama.cpp, Transformers, ExLlama, others GGUF, GPTQ, EXL2, AWQ Web-based local GUI OpenAI-compatible server Open source (AGPLv3)
Open WebUI Self-hosted web platform Feature-rich, self-hosted chat interface for teams, built primarily for Ollama and OpenAI-compatible backends Docker, self-hosted (any OS) Backend-agnostic (Ollama, OpenAI-compatible APIs) Depends on backend Browser-based, multi-user Consumes OpenAI-compatible APIs Open source (MIT)
AnythingLLM Self-hosted web platform All-in-one document/RAG workspace with zero-config retrieval and built-in agents (web search, SQL) Desktop (macOS/Windows/Linux) + Docker Backend-agnostic (Ollama, LocalAI, OpenAI-compatible) Depends on backend Native app or browser (Docker mode) Consumes and exposes OpenAI-compatible APIs Open source (MIT)
LibreChat Self-hosted web platform Enterprise-ready, open-source multi-model chat platform with agents, code execution, and persistent memory Docker, self-hosted Backend-agnostic Depends on backend Browser-based, multi-user Consumes multiple provider APIs, including local Open source (MIT)
PrivateGPT Self-hosted web platform Framework for fully private, offline document interaction, built for strict air-gapped deployments Local, Docker, cloud llama.cpp, LlamaIndex GGUF API-first; minimal bundled UI REST API Open source (Apache 2.0)
LocalAI Inference server Drop-in, OpenAI-API-compatible REST server for local inference; infrastructure component rather than a chat app Docker, Linux, macOS, Windows llama.cpp and others (ggml-based) GGUF and others None (API only) OpenAI-compatible REST API Open source (MIT)
llama.cpp Inference engine The foundational C/C++ inference engine underlying most GGUF-based local LLM tools Cross-platform, including low-resource hardware Native GGUF Minimal CLI/server; bundled web UI OpenAI-compatible server mode Open source (MIT)
vLLM Inference engine High-throughput serving engine optimized for concurrent, production-scale inference rather than single-user chat Linux, Docker Native (PagedAttention) Safetensors, GGUF (limited) None (API only) OpenAI-compatible REST API Open source (Apache 2.0)
koboldcpp Inference engine / chat app Single-binary llama.cpp fork popular for story-writing and roleplay use cases macOS, Windows, Linux llama.cpp fork GGUF Bundled lightweight web UI OpenAI-compatible + native API Open source (AGPLv3)

Notes

  • “Engine/Backend” refers to the underlying inference technology; several front-end tools (Open WebUI, AnythingLLM, LibreChat) are backend-agnostic and typically run against Ollama, LocalAI, or another OpenAI-compatible server rather than performing inference themselves.
  • Many details - especially license and platform details are subject to change; verify against each project’s official repository before citing as current.

That’s a lot of projects to consider and I’m sure there are many contenders missing. If you have personal experience with any of them, please share your experience and warnings/advocacy!

1 Like

Might not be too applicable to others, but my setup is:

  • Mac Studio M4 128GB of RAM
  • LM Studio with Qwen 3.8 9B and 27B
  • LM serves the models to Hermes Desktop

I use 9B for very quick chatting. Being agentic, Hermes slows things down so I might just use LM Studio. But if I want it to do something I go through Hermes. I have 27B for particularly complicated tasks.

Pretty simple setup but I am largely satisfied. All local. I am not using anything to to connect to it remotely. I have LM Studio with Qwen 9B on my MacBook Air M1.

This is great start! I’ll edit on desktop later.

I’ve also slowly been collecting Bring Your Own Key (BYOK) apps. This means you can plug in any provider include a local one.

This is not quite the same as Local AI, but it also doesn’t necessarily mean self hosting or other things: it’s bringing your own inference.