Liquid AI d1-3B: what it is and is it free to use?

Liquid AI d1-3B is a small open-weight model that doesn't write text. It answers yes/no, multiple-choice and score questions about a piece of text or an image, in one pass and in a few milliseconds. It's free to download and free for commercial use if your company makes under $10 million a year. Above that, the license doesn't cover you.

The d1 name first showed up on X on October 5. In a post from @liquidai, the company said its API-hosted d1 "now supports images, text or both as inputs." It said it had tested d1 against GPT-6.1 Sol and Claude Opus 5.5 on six real applications: d1 matched or beat GPT-6.1 Sol on four, cost 19x to 200x less than both, and answered faster every time. Those are Liquid's own numbers. Two days later, on October 7, Liquid put two d1 models on Hugging Face as open weights: d1-3B and an experimental d1-omni-600M.

This is the fourth "decision model" launch we've covered in about three weeks, after Microsoft-Decision-1, the OpenAI Decisions API and Perplexity's. Unlike those three, Liquid's can be downloaded and run on your own hardware.

What d1-3B does

You give the model a "state," meaning a support ticket, a form submission, a JSON record or a photo, along with a few named questions. Each question is one of three types: yes/no, a choice from options you define, or a score on a scale you define. It returns a probability for each possible answer. Since it never generates text, it can't ramble or make things up in prose. Asking several questions about the same input also costs little more than asking one.

Liquid's example is a support ticket that the model checks three ways at once: is this a refund request, which team should get it, and how urgent is it. The model card lists the jobs it was built for: routing and triage, moderation, intent classification, extraction checks, reranking, scoring another AI's output, agent guardrails and visual inspection. The card also says plainly that it's not a chat model.

Under the hood it's built on Liquid's LFM2.5-VL-3B vision-language model. It has 3.12 billion parameters and a 32,768-token context window.

How fast it is

Liquid published latency tables measured with NVIDIA. For one question on d1-3B:

• RTX 4090 gaming GPU: 8 ms
• Apple M5 Pro laptop: 30 ms
• Jetson Orin Nano, a small embedded board: 50 ms
• Three questions on the 4090: 21 ms

Liquid says d1-3B scores 48.57 on Decision Index 0.2.1, which it calls the best result for any decision model under 10 billion parameters. On its spread of seven public datasets (reading comprehension, toxicity, intent, medical QA and cross-lingual tasks), the mean is 82.9.

Liquid itself tells you to take those scores with some salt. On the October 8 ThursdAI podcast, Liquid's head of post-training Maxime Labonne said "you cannot really trust these benchmarks" for decision models yet. He also admitted the category is "a lot of rebranding." That's fair, since this is essentially a classifier you configure with plain-English questions. The model card shows d1-3B scoring below its own base model on some vision tests.

How to run it

The Hugging Face blog post gives a short Python setup. Install transformers 5.14 or newer along with torch, torchvision and pillow, then load the model with trust_remote_code=True. Call system_one() with your text, your questions and an optional image, or system_one_batch() to pack many requests together. The model card also lists 17 quantized versions that work with llama.cpp, Ollama and LM Studio, which puts it within reach of an ordinary Mac. If you'd rather try it before installing anything, Liquid has a demo Space on Hugging Face called System One Arcade.

Is d1-3B free?

Mostly yes, with one catch. Liquid's blog says you can "download, fine-tune, and deploy without restrictions." The license file on Hugging Face (LFM 1.0) is narrower. Commercial use only applies if your organization stays under $10 million in annual revenue, and any company over that line "is not licensed under this Agreement." Qualified nonprofits get an exemption, but only for non-commercial or research use. So d1 is open-weight, not open source in the Apache sense. Read the license before you build a product on it.

Our read

Most small businesses won't touch Python, so for them the price matters more than the download. Sorting incoming leads, flagging angry emails, checking that an intake form is complete, deciding whether a photo shows the damage a claim describes: those are all yes/no or pick-one questions. Plenty of shops either send them to a big chat model at chat-model prices or pay a person to handle them. If a model answers them in milliseconds on hardware you already own, and costs nothing under $10 million in revenue, you can afford to run it on every message that comes in.

Before you pick a model, write down the decision you're actually making. If you can't name the decisions in your inbox or intake process, our free process audit is a good place to find them.