Liquid AI Releases d1.2 Decision Models

The new models accept text, image and audio input and return calibrated probabilities over fixed answers without generating output tokens.

Updated on Nov. 17, 2026 in Artificial Intelligence

Isometric editorial illustration of a row of sealed grey server cabinets along a dim corridor, each with a small amber indicator light.
Liquid AI released d1.2 and d1.2-ultra decision models on Tuesday, deprecating its earlier d1 family and offering hosted-only access with no weights released. AI Illustration. Upload story photo >

Liquid AI released d1.2 and d1.2-ultra decision models, deprecating its earlier d1, d1.1 and d1-ultra models. The d1.2 family accepts text, images and audio and adds five new answer types on the same endpoint.

Why it matters

The models return calibrated probabilities over a fixed set of answers and generate zero output tokens, a design aimed at classification-style decisions rather than open-ended text generation. The release also shifts access toward a hosted-only API, with no model weights made available.

Liquid AI reports a 0.8-point calibration error for d1.2-ultra across 13 benchmarks in its own test, with d1.2 scoring 69 and d1.2-ultra scoring 75 on the Artificial Analysis Decision Index v2.1. Parameter counts remain undisclosed, and no independent verification of the calibration figure was provided.

The players

Liquid AI

Liquid AI is an artificial intelligence company that develops alternative model architectures and released the d1.2 and d1.2-ultra decision models.

The details

Both models are hosted on the Liquid API only, with no weights released; Liquid AI says it will open the weights "at some point." Requests accept up to 65,536 input tokens and 128 questions per call, and audio is accepted as 16-bit PCM WAV and billed at 10 tokens per second. New organizations receive $2 in free credit on Liquid Console.

Timeline

  1. November 17, 2026: Liquid AI released d1.2 and d1.2-ultra and deprecated d1, d1.1 and d1-ultra.

Roadmap

The release extends Liquid AI's d1 line into a d1.2 generation scored on the Artificial Analysis Decision Index v2.1, a benchmark used to compare decision-oriented models rather than general chat systems. It also follows a pattern of vendors moving specialized classifiers into hosted APIs rather than open weights, with Liquid AI stating only that weights will open "at some point."

Developers can call the models through the Liquid API at $0.02 per 1M input tokens for d1.2 or $0.042 for d1.2-ultra, with $2 of free credit for new organizations. Because the models generate no output tokens and cannot be self-hosted from released weights, teams relying on them depend on Liquid AI's hosted service and its pricing.

The takeaway

Decision models that return calibrated probabilities without generating text may suit classification workloads where output cost and predictability matter more than open-ended responses. Teams evaluating them should weigh the hosted-only access and undisclosed parameter counts alongside the benchmark scores.

Further reading

For broader context on how these systems are being built and deployed, explore our Artificial Intelligence coverage.

More information

Review the Liquid AI decisions API endpoint for request limits and access details.

Source note: This article includes information reported by MarkTechPost.