guides · featured

Best local LLM tools (2026)

Which local stack to install: Ollama, LM Studio, llama.cpp, Open WebUI. Table-first, linked to catalog pages. Grounded in GitHub and vendor sites fetched 2026-09-09.
by cache.directory updated Sep 9, 2026 CC-BY-4.0
← back to shelf

## what it does

tl;dr: Ollama if you want a CLI/API runtime. LM Studio Bionic if you want a desktop agent. llama.cpp if you want the engine. Open WebUI if you want a self-hosted chat UI on top of any of those. We did not benchmark models this session.

Quick pick

You wantInstallLicense (this fetch)GitHub stars (2026-09-09)
One CLI + REST on :11434, Docker, scriptsOllamaMIT180,524
Desktop agent (docs, skills, MCP, GUI)LM StudioProprietaryn/a (app is not the SDK repo)
GGUF engine, llama-server, every backendllama.cppMIT127,634
Browser UI, RAG, RBAC, on Ollama or OpenAI-shaped APIsOpen WebUICustom (branding clause)151,463

Head search term in our Keyword Planner dump is ollama (500k bucket, competition 1). LM Studio / llama.cpp / Open WebUI sit in the 50k bucket. This page is the collection those queries should land on.

Versions we fetched today

ToolCurrentDate on the release/changelog
Ollamav0.33.32 Sep 2026 (release)
LM StudioBionic 1.1.28 Sep 2026 (changelog)
llama.cppv0.4.04 Sep 2026 (release)
Open WebUIv0.11.3 (as of the catalog page’s fetch)31 Aug 2026

Ollama v0.34.0 was in pre-release when the Ollama catalog page was written; we are not treating pre-release as current.

How the pieces compose

llama.cpp          ← engine (GGUF, Metal/CUDA/Vulkan/…)
    ↑
Ollama             ← library + CLI + :11434
LM Studio runtime  ← MLX + llama.cpp packs inside the app
    ↑
Open WebUI         ← chat UI; points at Ollama, LM Studio, or any OpenAI-compatible URL

You do not pick one forever. Homelab default: Ollama (or llama-server) + Open WebUI. Laptop default: LM Studio, or Ollama + ollama launch claude.

By job

JobPickWhy
Backend for an app, CI, DockerOllamaHeadless, MIT, official install script + Docker image
Claude Code / Codex against local weightsOllama (ollama launch claude)Documented on ollama.com this session
Need MIT, will compile flags yourselfllama.cppllama-cli / llama-server; repo moved to ggml-org/llama.cpp
Documents, slides, skills, MCP in a GUILM Studio BionicAgent loop is the product; Linux landed in 1.1.2
Team chat UI, RAG, SSO, channelsOpen WebUISelf-hosted; Docker is the documented path
Same GGUF, no wrapperllama.cppThinnest layer

Pairing page: Ollama vs LM Studio.

Hardware (from the catalog pages, not a bake-off)

  • CPU-only works. llama.cpp is explicit about it. Small quantized models on Ollama list 4GB RAM as a floor.
  • Apple Silicon: both Ollama and LM Studio have MLX paths. LM Studio 1.1.0 changelog claims 2–2.75× faster prompt processing on M5 via MLX. Untested here.
  • NVIDIA: Ollama Docker, llama.cpp CUDA, Open WebUI :ollama CUDA image.
  • Do not quote tokens/sec we did not measure.

What this page is not

  • Not “best 7B model.” Weights move weekly; use Ollama’s library, the Hugging Face Hub, and GGUF repos, then come back for the runtime.
  • sandbox_test.verdict stays untested.

## embed this badge

cache ✓ in cache.directory
![cached](https://cache.directory/badge/best-local-llm.svg)