stick_ai not on sale yet

// a private coding assistant on a usb stick

Your code never leaves the stick.

Plug it in, open it, ask. Stellan reads your project, plans the change and waits for your yes before it touches a line. The AI runs on your own computer. There's no account, no cloud and no server to send anything to.

  • 73.2% HumanEval
  • 0 bytes sent anywhere
  • Mac · Win · Linux one stick
The stick_ai USB stick: a dark stick with a glowing violet edge, a small terminal face on its cap and stick_ai printed on its side
fig. 1 stick_ai, 32 GB · not actual size

// how it works

It asks before it acts.

  1. 01
    Plug in, open. The app for your computer is right on the stick. Nothing to install.
  2. 02
    Ask in plain words. Stellan explores the project you choose, and only that project.
  3. 03
    Say yes or no. Every edit and command shows the exact lines first. Your file stays as it was until you allow it.

// benchmarks

Better than GPT-4 and Claude 2 at coding. On your laptop.

On HumanEval, the classic coding test, Stellan-1 Lite scored higher than GPT-4 and Claude 2, OpenAI's and Anthropic's best models until November 2023. With 7.3 billion parameters, it's about 99.6% smaller than GPT-4's estimated 1.8 trillion.

stellan-1 lite73.27.3B params · your laptop
claude 271.2Anthropic's best, Jul–Nov 2023 · size unpublished
gpt-467.0OpenAI's best, Mar–Nov 2023 · ~1.8T params (est.)
gpt-3.548.1OpenAI, 2022
99.6%smaller than GPT-4 by its estimated size, and it scores higher
~5 sfor a follow-up step, down from ~95 s, because it remembers the conversation
4.1 GBLite's whole model. Runs on an ordinary computer with 8 GB of memory
How we measured, and the small print

Stellan-1 Lite: all 164 HumanEval problems, pass@1, greedy decoding (temperature 0), each answer run against the official hidden tests, on a 4-core CPU with the app's own settings: 120 of 164 passed.

Other models: scores as published by their makers: GPT-4 and GPT-3.5 in OpenAI's GPT-4 Technical Report (March 2023, 0-shot), Claude 2 in Anthropic's Claude 2 announcement (July 2023). GPT-4 was OpenAI's top model until GPT-4 Turbo (6 November 2023); Claude 2 was Anthropic's until Claude 2.1 (21 November 2023). Each company used its own prompts and setup, so treat this as a guide, not a strict head-to-head.

Model sizes: Stellan-1 Lite has 7.29 billion parameters, counted from its model file. OpenAI has never published GPT-4's size; ~1.8 trillion is the most widely cited estimate, from a July 2023 SemiAnalysis report, so "99.6% smaller" (7.29B ÷ 1.8T ≈ 0.4%) rests on that estimate. Anthropic has never published Claude 2's size and there's no reliable estimate, so we don't give one.

To be clear: newer cloud models from OpenAI and Anthropic score over 90% on this test (Anthropic reported 92.0% for Claude 3.5 Sonnet in 2024) and are far larger. stick_ai gives you 2023-flagship-level coding help, privately, offline, on your own machine. Stellan-1 Max's score is still being measured.

SPEC SHEETstick_ai · rev 1

Specifications

Models
Stellan-1 Lite (7.3B) · Stellan-1 Max (23.6B)
Runs on
macOS 13.5+ (Apple Silicon & Intel) · Windows 10/11 · Linux x64
Memory
Lite: 8 GB · Max: 16 GB minimum, 24 GB+ recommended
Internet
Not needed. Not used.
Stick
32 GB USB, exFAT, everything preinstalled
Approvals
Every edit, command and tool call
Subscription
None. Buy it once.

// what's on it

One stick, every computer.

The app for each computer sits at the top. Everything they share, including one copy of each model, lives in stellan/.

STICK_AI/
├── Stick AI.app     mac (apple silicon + intel)
├── Stick AI.exe     windows
├── stick-ai         linux
└── stellan/
    ├── models/
    │   ├── stellan-1-lite-q4_k_m.gguf
    │   └── stellan-1.0-q4_k_m.gguf
    ├── app/  bin/      the engine, per computer
    └── manifest.json   every file, checked

// launch price

Price

€15

per stick, VAT included · $17 outside the EU

+ shipping, calculated at checkout

  • The stick with the app, Lite and Max ready to go
  • Mac, Windows and Linux on the same stick
  • No subscription, no account

// not on sale yet

Ships from

EUFEB 2027orders & shipping
WORLDMAY 2027orders & shipping

Shipping is paid by the customer and added at checkout, based on where you live. Prices may change before launch.

Get one email when orders open

// questions

Questions.

When can I buy it?

It isn't on sale yet. Orders open in February 2027 with shipping across the European Union, and in May 2027 for the rest of the world.

How much is it?

€15 per stick in the EU, VAT included, and $17 in the rest of the world, plus shipping. It's a one-off price: no subscription.

Is it really better than GPT-4 and Claude 2?

On HumanEval, a well-known coding test, Stellan-1 Lite scored 73.2%. GPT-4 scored 67.0% and Claude 2 scored 71.2%, as published by OpenAI and Anthropic. Both were their company's best model until November 2023. Today's newest cloud models score higher still. What makes stick_ai different is that it runs privately, on your own computer.

Does it need the internet?

No. Stellan runs entirely on your computer. It only goes online to check for app updates, if you connect GitHub, or for a web search you allow (only the search words are sent, never your code). Everything else works offline.

My computer says it can't open the app. Is it safe?

Yes. stick_ai isn't signed with a paid Apple or Microsoft certificate, so your computer may ask once before opening it. Here's how to open it: it takes a few clicks and only happens the first time.

Does the AI get copied onto my computer?

No. The models stay on the stick. While the app is open they're read into your computer's memory, like playing a film straight off a USB drive. Nothing is installed.

Which computers does it work on?

Macs with Apple Silicon or Intel (macOS 13.5 or newer), Windows 10 and 11 (64-bit) and 64-bit Linux. Lite needs 8 GB of memory; Max needs at least 16 GB and runs best with 24 GB or more.

Do I pay for shipping?

Yes. Shipping is added at checkout and depends on where the stick is going.