最新模型
Liquid AI Releases d1.2 Decision Models, Adding Audio Input and Five Structured Task Types
On November 16, Liquid AI launched d1.2 and d1.2-ultra, adding audio input, document field extraction, image localization, and other capabilities. The new versions provide structured answers and probability information through an API; accuracy for the new tasks is still based mainly on the company’s internal testing.
Liquid AI released d1.2 and d1.2-ultra on November 16, expanding the decision models’ uses to audio analysis and a broader range of data extraction tasks. Both models are available through the Liquid API. The engineering focus of this update is to return classifications, localization results, and extracted data in specified types, reducing the need for applications to parse free-text answers. Official announcement
The five new task types cover reading text, numbers, and dates; extracting text spans; drawing bounding boxes around objects in images; estimating quantities with ranges; and multiple-choice classification. Audio can be the content being analyzed or carry a spoken question, such as asking the model to identify a language or determine whether a siren is audible in the background. The company says the new answer types use the existing endpoint and request format. Release notes
According to the documentation, decision models return results in a single call rather than generating text token by token. Extraction answers can include the source text location and found field; bounding boxes include coordinates and confidence information. Missing values have an explicit representation. For document processing systems, this can help incorporate “not found” into data validation, but engineers still need to distinguish missing content from parsing failures and calibrate decision thresholds using their own data. Technical documentation
The d1.2 page from third-party evaluator Artificial Analysis lists a Decision Index score of 69, latency of 150 milliseconds per task, and a cost of about $63 per million evaluated tasks. The index accounts for both answer correctness and probability quality, penalizing overconfident errors more heavily. The latency and cost figures apply to its evaluation workload and cannot be directly extrapolated to enterprise document or audio workflows. Evaluation and methodology
Official input prices are $0.02 and $0.042 per million tokens, with charges for input only. Before deployment, teams should validate Traditional Chinese fields, date formats, low-quality scans, and real-world recordings. The company also states that there are no corresponding public benchmarks yet for the five new task types; its reported results come from internal testing. Independent replication and whether confidence thresholds can reliably control errors on real-world data distributions are worth watching next. Pricing and limitations