Liquid AI d1.2 and d1.2-ultra: Why Automation Needs Decision Models Instead of ChatbotsNote · digest for November 19
Modular clay composition of geometric blocks with precise perforations and data structure
Note3 sections≈ 5 min

Liquid AI d1.2 and d1.2-ultra: Why Automation Needs Decision Models Instead of Chatbots

Why plug in a specialized model like Liquid AI d1.2 when ChatGPT or Claude can already output JSON from a prompt? Every chatbot is built for flowing conversation. Even under a strict schema, it generates text character by character, charges you for output tokens, and sometimes invents missing fields.

On November 16, 2026, Liquid AI released its second generation of decision models—d1.2 and d1.2-ultra. Instead of prose, they return structured outputs with confidence scores across text, photos, and audio.

Why chatbots make awkward gears and how decision models work

When a standard language model processes a form, it acts like a narrator: it reads the input and types an answer token by token. The d1 decision series works differently. In one forward pass through the network, it immediately outputs ready JSON with a decision and a calibrated probability.

That leads to unusual billing: you pay only for input tokens, and zero for output. A million input tokens cost $0.02 on d1.2 and $0.042 on d1.2-ultra, both with a 66K-token context window.

On the Artificial Analysis Decision Index v2.1, the larger d1.2-ultra scored 75 and took first place out of 25 models against a median of 66. The score combines six tests and penalizes confident mistakes. Smaller d1.2 scored 69.

A chatbot writes an essay and charges for every letter, while a decision model quietly checks the boxes.

Five new tasks and audio input: what d1.2 can do

Take a website support ticket: a customer fills out a form, attaches an invoice scan, and uploads a photo of a dented box. Here is how five new features in the Liquid AI release handle that stream.

First is typed extraction for text, number, and date. The model pulls an amount and date from a form or PDF along with a verification quote. If a field like VAT isn't on the invoice, it honestly returns not found. On CORD and SROIE receipts, accuracy reaches 98% for d1.2-ultra and 96% for d1.2.

Second is span detection. Say you need to pull every phone number, email, or customer question from a thread without manual search. The model returns each span and its exact text positions: on Nemotron PII personal data, Ultra hits 94% precision and 93% recall, while d1.2 gets 92% and 88%.

Third is object localization via box. The model spots a signature, logo, or defect on a photo and returns bounding-box coordinates around the target. Images are billed at 1.5 tokens per 32×32 pixel patch—a 1024×1024 frame uses 1,536 tokens—and follow-up questions on the same image cost only the question tokens. Both versions hit 94% on RefCOCO.

Fourth is numerical estimation with a range. When counting items on a shelf or people in a photo, instead of one rigid number it gives an interval, such as 12 with a range of 10 to 14. On COCO within plus or minus one object, Ultra reaches 93% and the base model reaches 92%.

Fifth is multi-label classification via choice with multiple: true. Customer emails rarely fit one bucket, so a single ticket gets multiple tags with individual probabilities: "shipping," "billing," and "urgent." On COCO multi-choice, d1.2 scores 90% and Ultra scores 87%.

Beyond text and photos, the models accept audio either as a clip to analyze or as a spoken question. On held-out sets from ESC-50 and FLEURS, they classify sounds and 43 spoken languages at 95–97% accuracy.

Confidence thresholds over blind trust, plus real limits

Regular models often claim 95% confidence and still get things wrong. Across 13 benchmarks totaling 22,505 questions, d1.2-ultra posted a calibration error of 0.8 points, and d1.2 posted 1.2 points. Rivals show wider gaps: 2.7 for Jev 1.14, 3.7 for Perplexity Decider v1.3, and 7.2 for GPT-6 Luna Decisions API. When Liquid AI's model outputs 80%, it is right roughly four times out of five.

In Make, n8n, or your own backend, that gives you a safety net via confidence thresholds. Tickets whose fields and tags like "shipping," "billing," and "urgent" clear your bar go straight to the CRM and the right team. Borderline cases below the bar route to a human operator, so your automation doesn't fail silently.

There are four limits to keep in mind. First, d1.2 doesn't write emails or code—you still need a regular language model beside it to draft replies. Second, unlike local Open d1 and LFM models that respond under 50 ms, these releases are closed and run only via Liquid AI's cloud API or OpenRouter, with parameter counts undisclosed.

Third, d1.2-ultra takes 218 ms per task on Artificial Analysis (vs. a 206 ms median) at an average cost of $133 per million tasks. Where speed is critical, d1.2 is the better pick at ~150 ms in tests and a 0.30-second median on OpenRouter. Finally, the 87–98% scores across the five new tasks come from Liquid AI's internal tests on held-out slices of public datasets, not an independent leaderboard.

Want a deep diveVotes help us choose what to break down in detail. This material has 0 votes.
Sources · 5