Multimodal decision model returning calibrated probabilities for text and image questions, with no generated tokens and a #1 Decision Index score.
Output Speed *
Intelligence Index *
Context Window *
Input price
Output price
Performance
Abstention and Calibration Behavior Drive d1-ultra's Ranking
Benchmarks
Independent Decision Index Results: Rank, Hallucination and Per-Task Latency
Output Speed
Intelligence Index
MMLU *
GPQA *
HLE *
LiveCodeBench *
Technical Specifications
What the model supports
Comparison
d1-ultra or Jev 1.14 for Hosted Decision Workloads?
d1-ultra and Jev 1.14 (TypeSafe) are both hosted decision models. Both return a probability for each answer to typed questions instead of generating text, and both bill input tokens only at the same per-token rate. The decision is therefore about accuracy and calibration, not cost. On Artificial Analysis Decision Index v1.0, d1-ultra scores 65 against 57 for Jev 1.14. In Liquid's own 112-evaluation suite it averages 76.7 versus 69.4, a vendor-reported result.
Choose d1-ultra when you want the stronger independent Decision Index result, calibration-focused training for high-stakes automation, or image-plus-text checks. Choose Jev 1.14 when it already sits in your stack and passes your own evaluation. With equal per-token pricing, switching only pays off if d1-ultra's gains show up on your data. Both are poor fits for tasks that need generated text. Shadow-test both on identical inputs before moving traffic.
Limitations & Trade-offs
Best-Fit Workloads
Where this model earns its place
Visual inspection against a spec
Liquid positions d1-ultra for hard visual cases, such as spotting a defect on a circuit board against its specification. It accepts text and images together and can return a choice with a probability per defect label. Liquid reports 77.9 on its picture category, a vendor figure. Use it for manufacturing quality checks that route low-confidence images to a human inspector. It won't generate inspection reports, so pair it with a text model for documentation.
Claim and listing verification
Liquid cites checking a damage claim against the customer's photo, and its demo compares product images with listing details to flag mismatches. d1-ultra's high non-hallucination rate suits this: it clears clear-cut cases and flags uncertain ones. Insurance intake and e-commerce catalog QA both benefit. Because each question re-bills every attached image, group questions deliberately and validate accuracy on your own claim and catalog data first.
Confidence-gated routing and triage
Liquid trained d1-ultra for calibration, reporting a 1.5-point mean gap between confidence and accuracy against 2.7 for Jev 1.14 (vendor chart). That makes its probabilities usable as thresholds: automate above a cutoff, escalate the rest. Typical uses are agent routers, cascade gates and support-ticket triage. One call can answer several typed questions about the same input. Calibrate thresholds on your own labeled data, because Liquid's calibration figures are not independently reproduced.
Multilingual moderation and guardrails
Liquid states support for 44 languages and lists moderation and guardrail checks among intended uses. In its own suite d1-ultra scores 80.2 across languages and 89.3 on robustness (vendor-reported). Use it as a pre- or post-filter on user content and agent actions, returning a policy-violation probability. Validate coverage for your specific languages first, since the 44-language claim is not independently broken down per language.