Liquid AI Releases d1.2 Omnimodal Decision Model, Tops Decision Index

Liquid AI releases d1.2, a zero-output-token omnimodal decision model that replaces text generation with calibrated probabilities.
Liquid AI launches the d1.2 and d1.2-ultra decision models, supporting omnimodal text, image, and audio input but generating no text—outputting calibrated probabilities directly. d1.2-ultra scores 75 on the Artificial Analysis Decision Index v2.1. The design bypasses token-by-token decoding, potentially offering cost advantages on batch decision tasks, and probability output naturally fits rule engines and downstream classifiers. Positioned as a decision component rather than a conversational assistant, it suits risk control, content moderation, and workflow branching. The trade-off is losing conversation, explanation, and multi-turn interaction. Its success depends on whether decision-specific evaluations gain industry consensus and whether enterprises are willing to restructure systems for probability output.
Liquid AI has released a new generation of decision models, d1.2 and d1.2-ultra. Unlike mainstream large models, this family generates no text at all—it outputs only calibrated probability values—yet it scored 75 on the Artificial Analysis Decision Index v2.1. It follows a technical path entirely different from conventional LLMs.
The Divide Between Decision Models and Generative Models
The defining feature of the d1.2 series is "omnimodal input, zero output tokens." The model can simultaneously read text, images, and audio, then directly produce a calibrated probability distribution for judgment, classification, or scoring.
This design confines the use case to decision-making rather than conversation. Conventional large models must first generate reasoning text and then extract a conclusion from it; d1.2 skips this step, and the output itself is the decision result. The benefit is a completely different latency and cost structure, and the calibrated probability values are better suited for direct embedding into business processes.
In terms of positioning, Liquid AI does not intend to compete head-on with large-model vendors in general-purpose conversation, but rather to target areas requiring stable, quantifiable judgment—such as risk control, content moderation, and branch decisions within automated workflows.
Calibrated probability means that the probability values output by the model are consistent with the actual event occurrence rate. For example, if the model assigns a confidence of 0.8 to 100 samples, then approximately 80 of them should belong to the positive class. Uncalibrated models (such as taking softmax output directly) tend to be overconfident, with inflated probability values, and using them directly for threshold decisions introduces systematic bias. Common calibration methods include temperature scaling, Platt scaling, and isotonic regression, with calibration parameters typically fitted on a validation set. For decision models, calibration quality matters more than accuracy, because downstream rule engines and risk-control thresholds depend directly on the absolute values of probabilities, not merely their ranking.
Performance Positioning of d1.2-ultra
According to a MarkTechPost report, d1.2-ultra scored 75 on the Artificial Analysis Decision Index v2.1. This index specifically measures a model's performance on decision-type tasks rather than language generation quality.
It should be noted that the original material did not disclose the specific ranking position of this score on the leaderboard, comparison baselines, or evaluation details, so "Top the AA Decision Index" should be understood as the series holding a leading position on this decision-specific index. Information such as the differences between the two versions, parameter counts, context length, and inference cost has also not been made public.
For developers, what is more noteworthy is the evaluation dimension itself: when a model no longer proves its ability through text output, the traditional benchmark system needs to be adjusted accordingly, and the value of specialized metrics such as the decision index will rise.
Artificial Analysis Decision Index is a specialized evaluation index launched by Artificial Analysis. It belongs to the same evaluation system as the company's better-known LLM overall leaderboard, but focuses on decision-type tasks rather than language generation quality. Such decision indices typically consist of a set of judgment, classification, ranking, or scoring tasks, weighted and aggregated by task type into a single score. Note that specialized indices and general benchmarks (such as MMLU and GPQA) cannot be directly compared across the board—the former measures a model's reliability as a decision component, while the latter measures knowledge and reasoning ability. When the model's output form changes from text to probability, the traditional evaluation paradigm based on scoring generated text becomes invalid, which is why metrics such as the decision index are attracting attention.
What Zero Output Tokens Mean
"Zero output tokens" is the most counterintuitive design of this model family. The bulk of a generative model's cost often lies in the decoding stage—the longer the output, the higher the cost. d1.2 outputs probabilities directly, bypassing the token-by-token generation process, which may bring significant cost advantages on batch decision tasks.
The other implication lies in integration. Probability output naturally fits rule engines, threshold decisions, and downstream classifiers, without the need for an additional result-parsing layer. For enterprise systems requiring interpretability and auditability, a well-calibrated probability value is easier to incorporate into existing processes than a passage of natural-language reasoning.
The trade-off is reduced flexibility: the model cannot answer questions, explain reasons, or engage in multi-turn interaction. It is only suitable for use as a decision component.
Decoding stage is the main source of cost for autoregressive generative models: the model must sample or take the argmax token by token, requiring a forward computation at each step, and the output length directly determines latency and compute overhead. The zero-output-token design means that after encoding the full multimodal input, the model maps directly to a probability distribution through an output head, completely skipping the autoregressive decoding loop. This makes inference cost closer to "a single forward pass" rather than the cumulative overhead of "input length + output length." In batch decision scenarios (such as scoring tens of thousands of items one by one), this difference in cost structure is amplified, but it also means the model cannot perform chain-of-thought reasoning, limiting the depth of complex judgments.
Application Space of Omnimodal Decision-Making
Simultaneous input across text, image, and audio modalities allows d1.2 to handle judgment tasks that single-modality models struggle to cover. For example, combining user speech and textual description for intent determination, or performing quality-inspection grading based on images and accompanying instructions.
Such tasks are characterized by heterogeneous input and simple output—exactly the strength of decision models. By contrast, having a general-purpose large model handle the same tasks usually requires multiple rounds of prompting and post-processing, resulting in a longer pipeline and higher uncertainty.
However, the source material did not provide specific customer cases or deployment scenarios; the above inferences are based on the model's capability characteristics, and actual results still await more public evaluations and user feedback.
A Differentiated Path Worth Watching
Liquid AI's choice reveals a trend: model capability differentiation is occurring not only in parameter scale but also in output form. While most vendors compete around generation quality, converging the model into a "pure decision-maker" is a pragmatic product strategy.
Whether it succeeds depends on two things: whether specialized evaluations such as the decision index can form an industry consensus, and whether enterprises are willing to restructure their existing systems for probability output. d1.2 has provided the technical answer; the market answer has yet to arrive.
Related articles

Gemini 4 Neon Leak: A Potential Strong Contender Against Opus 5.5
A Hacker News thread suggests Gemini 4 Neon looks like an Opus 5.5 competitor. This article unpacks the product positioning behind this community signal, shifts in the top-model competitive landscape, and what it means for developers' model selection strategy.

Claude Fable 5.2's Self-Authored Short Film 'Fable' Sparks Discussion
A short film 'Fable,' credited to Claude Fable 5.2 for script, visuals, and directing, appears on Hacker News. This article analyzes the significance of full-pipeline AI creation, community reactions, and what it means for content creators.

Anthropic IPO Pricing: A Signal of AI Giants Going Public
Anthropic's IPO pricing draws attention. This article analyzes its potential impact on AI industry valuation, commercialization paths, and the competitive landscape, and highlights public-market expectations and challenges for AI companies.