The new v0.6.0 release packs a lot of good stuff:

• Clef support (text + vision)
• High-quality support for Qwen3.8-Flash-Next
• Massive performance improvement with Metal
• New llama_batch_ext API

Also the website got a nice refresh: https://t.co/9akc1jmG9t

Llama · Your AI, on your computer

Source: https://llama.app/ By thellama.cppteam and Hugging FaceLlama is a tiny menu bar app that runs the latest open models on your Mac.Llama is a system tray app that runs the latest open models on your PC.Llama is an app for Mac and Windows that runs the latest open models on your computer.Chat with them, or use them in your other apps.

Download for MacDownload for Windows 11View on GitHubView on GitHub

1 MB download ·Free · Open source · Works offline

Like OpenAI, but on yourMaccomputer

OpenAI has ChatGPT to chat with and an API to build on. Llama gives you both, running on your own computer, with models you choose.

For you

Chat· like ChatGPT

Click “Open chat” in the menu. Nothing to set up.

For your apps

API· like the OpenAI API

EditorsOpenAI-compatible

Chat appsOpenAI-compatible

Your own codeAny OpenAI SDK

Point any app that works with OpenAI atlocalhost:9931/v1

Your models· download once, use everywhere

Qwen3.827Bgpt-oss20Bgemma-4E4B

Nothing to set up

Running AI locally used to mean reading forum threads about settings. Llama checks yourMaccomputerand sets up each model to run well on it, based on how llama.cpp actually works. You only pick which model to talk to.

Know what you’re doing? Every llama.cpp setting is still there, in one plain-text file.

Qwen3.827BChosen by Llama, for yourMaccomputer

• QuantizationHighest precision that fits in memory
• Context sizeOnly sizes that fit in memory
• Image inputOn for models that understand images
• Speculative decodingOn for models that support it
• Batch sizeLarger on Macs with 32 GB or more

Light enough to forget it’s there

Local AI has a reputation for eating disk space and memory. Llama keeps both to a minimum.

4 MB app

A native Mac app, and just a 1 MB download — smaller than a photo.

Each model stored once

Kept in the Hugging Face cache, shared with llama.cpp and other tools.

Nothing loaded when idle

Models load when something asks for one and unload after 5 minutes idle.

A great model for everyMaccomputer

Llama suggests one that fits when you open it. Here’s where to start.

16 GBMacof memoryor moreGemma 4Everyday questions, writing, and images24 GBMacof memoryor moreGPT-OSSStep-by-step reasoning from OpenAI32 GBMacof memoryor moreQwen 3.8Strong at code, reasoning, and long documents

Browse models

Build on Llama

Your apps get a local AI API they can all share, and Llama takes care of the engine and the models.

OpenAI-compatible API

If your code works with OpenAI, it works with Llama. Change the base URL and keep everything else — no API keys, no usage bills.

• OpenAI- and Anthropic-compatible endpoints
• Streaming, tool calling, structured output, vision
• Already use llama.cpp? Your models show up automatically
• Thellamacommand too:llama cli,llama serve, and more
• Reach it from your other devices over Tailscale

API referenceCommand line

$curlhttp://localhost:9931/v1/chat/completions\-H“Content-Type: application/json“\-d’{ “model”: “ggml-org/gpt-oss-20b-GGUF:MXFP4”, “messages”: [{“role”: “user”, “content”: “Hi!”}] }’

Nothing to bundle

Your app talks to Llama over the API, and a one-click link installs the model it needs. No engine to ship, no gigabytes in your download — and your users keep one copy of each model for all their apps.

Without Llama57 GB on diskChat app

EngineModel19 GB

Coding agent

EngineModel19 GB

Your app

EngineModel19 GB

With Llama19 GB on diskChat appCoding agentYour app

Llama

EngineModel19 GB

Local AI starts here

Free, open source, and yours to keep.

Just want the command line?Install the command line:

curl \-LsSf https://llama\.app/install\.sh \| sh

irm https://llama\.app/install\.ps1 \| iex