Decision Tree, a little bonsai with a decision model inside

Decision Tree (opens in a new tab) is a small bonsai with a private decision model inside. You would ask it quick questions about text or photos on your home network, and it would answer with how sure it is. We added its page to offlinebase.com on October 7, 2026.

Where it stands

As of October 14, 2026, Decision Tree is marked "Coming soon" and "In development. Not available to order." It has no price and no date. The store (opens in a new tab) lists only Base Drive and Base Router.

The page closes with our plan for news about it: "We'll write when there's real hardware, a price, and a date. We'd rather under-promise." A note at the bottom of the page says its specifications, prices and dates "are targets, not commitments, and may change before launch." Read everything below as planned.

A model that picks answers

A chatbot writes its reply word by word. The model inside Decision Tree reads what you send once and picks from the answers you allow, so it can answer quickly.

We plan four kinds of question:

The examples show who might ask: a shop's website, an inbox app, a family member's phone or a coding agent on the same network.

The page puts it plainly: "Every answer says how sure it is. It can be wrong, so that number matters." A 91% yes and a 55% yes deserve different handling. Once the planned API exists, a builder could let an app act alone on confident answers and send the unsure ones to a person.

Why it is a tree

One answer can lead to the next. Our example follows one email ("I was charged twice this month") through three questions:

1. Needs a reply? Yes, 93%. A no (7%) would have sent it to the archive.
2. Which team? Billing, 88%. Support got 9% and Sales 3%.
3. How urgent? 3 of 4, 71% sure.

Chain small answers like these and you get a decision tree. In our drawing, thicker branches mean higher confidence, so the likely path reads as a trunk and the long shots read as twigs. In the planned API, one call can ask all three questions.

What the yellow light means

The canopy is designed to glow warm yellow whenever anyone uses it. In the page's words: "That's all the light means."

We are keeping the light to that single job. A glance at the shelf would tell you someone on your network just asked the tree something. The answer, with its number, would go only to the app or device that asked.

Planned hardware

There is no finished hardware yet, so we have no measured speed or power draw to share.

It would read only what other devices on your network send it: words or a photo.

The model inside

The planned model is d1-omni-600M (opens in a new tab) from Liquid AI. Its model card lists 587 million parameters and says it is built on LFM2.5-Encoder-350M. It returns typed decisions (a yes or no, a choice among named options, or a score) with zero output tokens (the word pieces a chatbot types out one at a time). It reads the answer straight from the probabilities the model gives each allowed option.

Liquid calls it "an early research release" that is under active development. Liquid opened its weights on October 7, 2026, in a post on its blog (opens in a new tab), the same day our page went up.

The model card names three question types: yes or no, choice (what we call pick one) and score. Our fourth type, pick any, is a set of yes-or-no questions about one photo ("Dog? Beach? Snow? Night?"), answered together in one read.

The model can also take short audio clips, up to 30 seconds of speech. Decision Tree is planned for text and images only.

Liquid also makes a larger sibling, d1-3B. Our Jetson fleet's router has used d1-3B since October 12, 2026, to decide whether a chat stays on our boards or goes to a cloud model. How a decision model picks between our Jetsons and the cloud covers that work. Decision Tree is meant to put the smaller model in a box on your own shelf.

For builders

We plan a local API on your home network, behind an API key, "with the same request shape as other decision-model APIs." The example request sends one piece of text and three named questions in a single call:

{
  "state": "I was charged twice this month.",
  "questions": {
    "needs_reply": { "type": "noul", "instructions": "Does this email need a reply from a person?" },
    "team": {
      "type": "choice",
      "instructions": "Which team should handle it?",
      "criteria": { "billing": "payments, charges, refunds", "support": "using the product", "sales": "buying or upgrading" }
    },
    "urgency": {
      "type": "score",
      "instructions": "How urgent is this, 1 to 4?",
      "criteria": ["can wait", "this week", "today", "right now"]
    }
  }
}

The example response gives an answer and a set of probabilities for each question, plus a separate confidence value on the pick-one and score questions.

Getting updates

Decision Tree now sits in our Products menu next to Base Drive and Base Router. Our overview, What Offline Base is building, and what you can use today, covers the rest of the lineup. The "Get updates" button on the Decision Tree page points to our newsletter (opens in a new tab), where news about Decision Tree will appear first.