Wink Pings

Liquid AI's New Models Skip Generating Full Answers, Output Only Calibrated Probabilities

Liquid AI d1.2 and d1.2-ultra support text, image and audio inputs, directly output calibrated probabilities with zero output tokens. The ultra variant scores 75 on the AA Decision Index v2.1, marking decision models as a category that is starting to meet formal workflow access thresholds.

Liquid AI has released its d1.2 and d1.2-ultra. Both decision models can process text, image and audio inputs, return calibrated probabilities, and produce zero output tokens. The d1.2-ultra scores 75 on the Artificial Analysis Decision Index v2.1.

This type of model is used very differently from chat models. Chat models turn inputs into a paragraph of response, while decision models convert inputs into a comparable, standardized score. For developers, the score can be directly connected to preset thresholds, rule engines and cost calculation pipelines. For regular users, you can think of it as a referee holding up a judging card: it only delivers the judgment result, no extra comments. Zero output tokens does not equal zero computation — the model still fully processes the input, but no text is generated on the output side, which changes the latency performance and billing structure. The exact amount of cost saved depends on the actual inference price and throughput of specific use cases.

Why is this worth noting? A large number of tasks in enterprise scenarios do not require a polished, fancy explanation. Content reviews that decide whether to escalate to human agents, support tickets that decide which queue to route to, images that need compliance checks, audio snippets that need anomaly detection — all of these problems boil down to several probability values. Natural language responses sound flexible, but they require extra parsing work when plugged into automated workflows. The value of calibrated probabilities lies in their adjustable thresholds and calculable misjudgment costs. Models that do not generate redundant text are actually more practical for real production work.

Liquid AI's d1.2-ultra gets a 75 score on AA Decision Index v2.1. This score comes from a third-party evaluation. The released material does not include the full leaderboard and comparison models, so the standalone number does not prove it is suitable for all scenarios. The real performance of multimodal decision models also depends on domain-specific data, probability calibration quality and latency. Key points developers should pay attention to include: whether the probability output is stable, whether the threshold still works after being migrated to actual business, and whether the cost of audio and image inputs is within budget.

Original author Asif Razzaq, source MarkTechPost: Liquid AI Releases d1.2 and d1.2-ultra.

发布时间: 2026-11-17 17:20