Liquid AI releases Open d1, open-weight decision models that answer in one pass without generating text

Why it matters

【Key points】

Liquid AI released Open d1 on October 7, 2026: two open-weight decision models on Hugging Face. They do not generate text; they read a state and typed questions (yes/no, a pick from named options, or a score) and return answers with probabilities in one forward pass, with zero output tokens.

d1-3B (3.12B parameters, built on LFM2.5-VL-3B) takes text and images. d1-omni-600M (587M parameters), which Liquid calls experimental, takes text with images or with up to 30 seconds of audio.

【Verifiable】

Primary: https://www.liquid.ai/blog/d1-open and the Hugging Face cards for LiquidAI/d1-3B and d1-omni-600M.

On the Decision Index v0.2.1, Liquid scored d1-3B at 48.57, above Decider 35B-A3B (47.11) and below Winnow-12B (50.02), using the official scorer, "not a leaderboard submission". d1-3B answers one question in 8 ms on an RTX 4090 and 50 ms on a Jetson Orin Nano. llama.cpp merged support for both models by early October 8 UTC.

Both use the LFM Open License v1.0: commercial rights hold only while the licensee stays under $10 million in annual revenue. The blog's "deploy without restrictions" is broader than that.

d1-3B's card promises "calibrated, typed answers"; d1-omni-600M's card says only its text answers are calibrated.

【Our read】

Decision models are moving from cloud APIs to devices. OpenAI's Decisions API is in public beta and TypeSafe's Jev opened in early access, both hosted; Liquid ships weights that run on a Jetson.

No text means control: every answer is a typed value with a probability, easy to log, threshold and audit. The cost is that the model only picks among options you define and cannot explain itself.

The licence is not OSI open source; larger companies need a separate deal.

Calibration and speed are self-measured. Independent checks matter more than for chat models, because downstream systems will act on these probabilities directly.