Liquid AI has released d1.2 and d1.2-ultra, omnimodal decision models that process text, images, and audio via a hosted API, returning calibrated probabilities over fixed answer sets without generating output tokens. The models support up to 65,536 input tokens per question and 128 questions per call, with audio encoded as 16-bit PCM WAV at 10 tokens per second. On the Artificial Analysis Decision Index v2.1, d1.2-ultra achieves the top score of 75, surpassing Drex Max v1.1 (73), SubQ Snap 1.2 (73), and GLiDE v1.2 (73), while d1.2 scores 69. Calibration error on Liquid AI's internal test across 13 benchmarks and 22,505 questions is 0.8 points for d1.2-ultra and 1.2 for d1.2, significantly lower than Perplexity Decider v1.3 (3.7) and GPT-6 Luna (7.2). The release introduces five new typed answer types: reading values with provenance, extracting spans with positions, locating objects via bounding boxes, estimating numbers with ranges, and multi-label classification with per-label probabilities. Accuracy on held-out datasets shows d1.2-ultra reaching 98% on receipt value extraction (CORD, SROIE), 94%/93% precision/recall on PII detection (Nemotron PII), 94% on object localization (RefCOCO), 93% on object counting (COCO), and 87% on multi-label picking (COCO), where d1.2 slightly outperforms at 90%. Pricing is $0.02 per 1M input tokens for d1.2 and $0.042 for d1.2-ultra; images cost 1.5 tokens per 32×32 patch. Model weights and parameter counts remain undisclosed, with Liquid AI stating they will be opened "at some point."

Read original