Liquid AI releases d1, a decision model that returns probabilities instead of text, adds vision and opens its weights

• Hosted model: d1 on the Liquid API. Vision requires the paid d1; the free tier (d1:free) is text only (Liquid Docs).
• Question types: Noul (yes/no, returns a probability from 0 to 1), Choice (a probability for each label) and Score (a position on a scale you define). Choice and Score also return a confidence value (Liquid Docs).
• Image limits: up to 8 images per request (JPEG, PNG, WebP or GIF), a 4.5 MB request cap, base64 only, and no remote image URLs (Liquid Docs; video).
• Vercel and OpenRouter: text only for now (Liquid AI blog).
• Open weights: d1-3B has 3.12B parameters and d1-omni-600M has 587M (OrcaRouter).

• How it answers: d1 reads the input and every question in one forward pass and returns probabilities, not generated text. Several questions about the same input fit in one request (Liquid AI blog; video).
• Architecture: the hosted model's backbone and size are not disclosed. For the open models, OrcaRouter describes d1-3B as decoder-based and d1-omni as encoder-based. HyperAI says both are built on Liquid Foundation Models (OrcaRouter; HyperAI).
• Image pricing: images are billed at 1.5 input tokens per 32×32-pixel patch, capped at 10,000 patches per request (Liquid Docs).
• Implementation: the docs show a three-way rule: block above 0.8, allow below 0.2, send anything in between to a person. A Noul result of 0.5 means "unsure", not "medium" (video).

• Speed and cost: 200–300 ms per text decision, and 19x to 200x cheaper than the frontier comparators (Liquid AI blog). The blog gives no per-task cost or latency figures.
• Head-to-head: matched or beat GPT-6.1 Sol on four of six applications. The two losses aren't named and per-task scores aren't published (Liquid AI blog).
• Visual inspection: 85–97% accuracy separating good from defective parts on four VisA product lines, with no task-specific training (AlphaSignal search listing).
• Agent and game demos: context compaction removed 52% of tokens while keeping the outputs the task needed; Tetris improved from 70 to 81 cleared lines with screenshots added; Wordle was solved 12 of 12 times in 3.8 guesses on average (video).
• Open models: d1-3B averages 74.1 across 11 public image benchmarks, and Liquid AI reports 8 ms per question on an RTX 4090. These are vendor numbers and have not been independently reproduced (OrcaRouter).

d1 sits in a line of classifier-style uses of language models: scoring candidate answers by likelihood instead of generating them, rerankers, and guard models. Commentators describe it as part of a move toward "zero-token" decision components inside agent pipelines (TechJack Solutions). Liquid AI's pitch matches this: sorting, routing, triage, moderation and approving agent actions. It explicitly steers free-form writing, multi-step reasoning, summaries and code generation to ordinary LLMs (video).

• Benchmark setup: each application was run once per model. Six of the 15 code questions and two of the four compaction sessions were written after d1's pipeline was already set. In visual inspection, every model saw a reference good part next to the part being checked (Liquid AI blog; video).
• Comparison handicap: the chat models answered in JSON at their default reasoning settings, which may not reflect how well they can do with tuning.
• Hands-on results: MindStudio's local test of d1-omni-600M found 51% confidence on an inferential vision question, and audio classifications that seemed to follow the speaker's delivery more than the meaning (MindStudio).
• Headline framing: the "200x" in the headline is the top of Liquid AI's 19x–200x range.

• Audit first: look for LLM calls that only produce a label, a flag or a score. Liquid AI's docs include an agent prompt for finding them (video).
• Pricing: output tokens are free, so cost depends entirely on input size. Asking several questions in one request reduces input spend.
• Use the probabilities: route borderline cases to a person or a larger model instead of trusting only the top answer.
• Local deployment: the open weights make on-device inspection and moderation possible, but none of the vendor numbers have been reproduced, so test d1 on your own labeled data first.