Private AI that runs on your Mac. In Privacy Mode, Local runs every request on your machine, without an account, telemetry, or per-token fees.
Download Local for macOS · Latest release notes
Requires macOS 14 or later and a Mac with Apple silicon. Windows and Linux versions are coming soon.
Released 2 September 2026. This release adds richer streaming Markdown and LaTeX rendering, tables that adapt to the available space, and a live model catalogue with an offline cache. It also includes fixes and polish across chat, meetings, and first launch.
Read the Local 0.1.1 release notes
Local is signed and notarized for macOS. It installs and manages BaseRT for you, checks for updates, and notifies you when a new version is available.
- Download the latest release.
- Open the downloaded DMG and install Local.
- Launch Local. It will detect your chip and memory and recommend a suitable setup.
- Download a recommended model and start chatting, coding, or capturing meetings.
Model downloads use additional disk space; the required amount depends on the model you choose.
- Chat with your files: ask questions, draft content, and summarise documents such as PDFs.
- Work on code: give Local access to a project folder so it can read, edit, and run the code in place without uploading the codebase.
- Capture meetings: record and transcribe meetings on-device, with speaker labels, summaries, decisions, and follow-up notes.
- Use memory you control: view and edit the facts Local remembers about you. Sensitive memories stay on your Mac.
- Track local usage: see generated tokens, speed by model, privacy statistics, and the estimated cost avoided compared with cloud APIs.
Local detects your chip and available memory, recommends models that fit, and sets itself up automatically.
Let Local read, edit, and run code inside a project folder you choose.
Record and transcribe meetings on-device, then create summaries, decision logs, and follow-up emails.
With Privacy Mode enabled, requests and memories remain on your device. After downloading Local and a model, you can use Local without an internet connection. Its data is stored locally on your Mac.
Cloud models are optional. You can add your own API key for OpenAI, Anthropic, or OpenRouter, or connect to a regional provider when a task needs a frontier model. Cloud requests are disabled while Privacy Mode is on.
The amount of unified memory determines which models your Mac can run; faster chips generate responses more quickly.
| Unified memory | Typical hardware | Best suited to |
|---|---|---|
| 8–16 GB | MacBook Air and base MacBook Pro | Chat with fast, compact models |
| 24 GB | Pro-chip MacBook Pro | Chat with capable mid-size models and everyday agentic coding |
| 32–128 GB | Max-chip Macs, Mac Studio, and desktops | Larger open models and full agentic coding |
Local supports open models from families including Qwen, Llama, Gemma, Mistral, Phi, DeepSeek, and Whisper, with more available in the app.
It includes BaseRT, Base Compute's inference engine. BaseRT is compiled for your Apple silicon on first launch and can run models up to 5.4× faster than the engines shipped by other local apps.
Office Mode lets multiple laptops use a model hosted on the most capable machine in your office. A Mac Studio, AMD Strix Halo system, NVIDIA DGX Spark, or on-premises server can do the inference while teammates connect from Local on their Macs.
Learn more at basecompute.co/local or join the Base Compute Discord.