INDEPENDENT MODEL · RESEARCH PREVIEW 0.7

An AI
of your own.

Original weights. Local inference.
A small beginning for a more personal AI.

Free preview · Apple Silicon · Internet needed for first setup

01 / WEIGHTSTrained from scratch
02 / ENGINEApple MLX
03 / CONTEXT1,024 tokens
04 / ACCESSYour Mac. Your copy.

A CLEAR STARTING POINT

A real model.
Still a research project.

Tars uses a project-owned transformer and tokenizer, trained from random initialization on a Mac mini. It runs the original model directly, with no pretrained-model or cloud fallback.

This preview is small and experimental. It can produce incorrect or incoherent replies. Autonomous tools are disabled, and personality controls are not yet reliably followed.

RECORDED LOCAL OUTPUTSAME RELEASED WEIGHTS

Write a Python function named add_numbers(a, b) that will return the sum of a and b. Output only Python code.

def add_numbers(a, b):
    return a + b
Verified against positive, signed and zero inputs.Example output, not live web chat.

MEASURED, NOT PROMISED

Know where
the model stands.

Small reserved task families for the released checkpoint. These scores do not establish general assistant or software engineering ability.

Basic Python functions restricted execution tests
32 / 32
Fresh repair math verified working required
3 / 40
Fresh repair reasoning separate generated problems
10 / 40
Fresh repair tool tasks experimental test workrooms
15 / 40

A LOCAL FIRST START

Set up once.
Then go offline.

  1. 01

    Download & unzip

    Keep the Tars AI folder somewhere writable on your Apple Silicon Mac.

  2. 02

    Open Setup.command

    Internet is needed once to install a private runtime and download 113 MB of verified original weights.

  3. 03

    Open Start.command

    Use Tars in your browser at localhost:8421. Keep the Terminal window open while it runs.

Developer preview: not signed or notarized. macOS may require approval in Privacy & Security. Tested on Apple M4, 16 GB memory, macOS 26.2. Intel Macs and iPhone-only inference are not supported.

BEFORE YOU BEGIN

A few things
to know.

Does it really run without internet?

Yes, after first setup. Generation runs on your Mac using the downloaded model. First setup and downloading future updates require internet. No account or online license check is required for this preview.

Can I use it from my iPhone?

Open Phone.command on the Mac, then use the local network address and pairing flow shown in the app. The Mac stays the brain; the iPhone is a remote on the same trusted network. Away-from-home access and a native iPhone app are not included.

Is this ready for autonomous work?

No. Earlier tool tests found unreliable multi-step behavior, so production tool execution remains disabled. This preview is for exploring an original model, not delegating important tasks.

What happens to my conversations?

The downloaded app stores its chat history locally in its own Tars AI folder. Your prompts are processed on your Mac and are not sent to a hosted model. The download contains no chats, credentials or workroom files from the developer’s Mac.

Is there a subscription?

The research preview is free and has no subscription expiration. Our optional $5/month supporter plan supports original-model research and future releases when available. No release schedule or advanced capability is guaranteed. See supporter details.

How do I verify the download?

The release metadata below includes the ZIP and model SHA-256 hashes. Setup checks every model part and the complete original weights before installation.

View checksums ↗

THE FIRST CHAPTER

Make room
for your own AI.

A free, experimental preview. Yours to explore locally.

Download for Mac

Apple Silicon · First setup needs internet · No account required