Liquid d1: Hosted Decisions with Channel-Specific Vision
Liquid d1 returns bounded choices, yes/no probabilities and rubric scores. Image input is available through the paid native API, with different limits on free and partner routes.
Checked October 6, 2026 · Documentation review, no hands-on product test
Choose the endpoint before the model name
Liquid presents d1 as a hosted decision model with Choice, Noul and Score outputs. Its documentation and announcement distinguish the paid native API's visual input from the text-only d1:free, Vercel and OpenRouter routes at this check. Planned channel support is not current support. Official announcement.
Use it for a defined decision, such as assigning an incoming request to a known queue. A probability is a signal for application logic, not permission to execute an action.
Image requests and billing are specific
The native /decisions/v1/systemone path accepts embedded base64 images, not remote image URLs. Current documentation permits up to eight images in a JSON body below 4.5 MB. Image examples use direct HTTP rather than implying every text SDK supports them. Decision and image request documentation.
Each question is billed for its text and all supplied images again. The documented 1024-square image costs 1,536 input tokens per question: two questions therefore add 3,072 image tokens, plus text. No generated output tokens does not mean a free request. Verify the account tariff and quota before calling. Billing rules.
Hosted access is not an open-weight release
The announcement describes available API access and future open-weight plans separately. We did not verify a released d1 checkpoint or assign an open-source license to the hosted service. Its single-run application experiments are vendor evidence, not a general speed, price or accuracy guarantee. Availability and experiment scope.
Example: triage a screenshot without inventing certainty
Define a small set of UI states, including “needs review.” Use the actual image-capable endpoint; validate that the image arrived before asking the decision question. A missing or ambiguous screenshot should stay unresolved, not become a success label.
For evaluation, keep labeled examples outside the tuning set and record endpoint, model, input size and questions. Decide fallback rules using observed errors rather than a universal probability cutoff. Compare hosted and local alternatives in the Decision Models selection guide.
Evidence and next step
Sources are linked beside the claims they support. Workflow suggestions are editorial examples, not completed tests; no account, installation, paid generation or benchmark was used for this profile.
Start with the official product or project and confirm the exact access, version and terms needed for your task. How we review.