decision-machine-1

Not yet independently verified. Dated independently only by the product's Product Hunt launch listing (21 September 2026), which is normally submitted by the maker; no press coverage was found. The description of the underlying models is CloudRaker's own. We will update this when it can be confirmed, and remove this note.

decision-machine-1 is the model behind Milliseconds.ai, an API from CloudRaker Inc. of Montréal for small, typed decisions on text and images. Rather than writing prose, it returns structured results: yes/no answers with a probability, a label with the full score distribution, a rating on a scale, answer spans with character offsets, fields matching a JSON Schema, named entities, and checks of a value against the source text. The API launched on 19 September 2026, the date of CloudRaker's own blog posts and of an engineering write-up by Baptiste Laget; it grew out of models CloudRaker already used in its Paperwork document API.

CloudRaker says the name refers to the product surface rather than a single newly trained model. Its runner uses a small instruction-following model, Qwen2.5-1.5B-Instruct, for constrained choices and GLiNER2.5 for span extraction. No parameter count or context window is published; each text may be up to 20,000 characters, and a request may batch up to 32 texts. Images (JPEG, PNG or WebP, up to 5 MB) are billed as a fixed number of tokens by detail level and are not stored. Labels and instructions should be written in English, though the input text may be in other languages. An OpenAI-compatible endpoint supports JSON-schema extraction and tool calls but not plain chat.

Production use costs $0.04 per million input tokens, with no charge for output tokens; test keys include 125 million free input tokens a month. The vendor quotes model time of about 90 ms for an image decision at medium detail.

CloudRaker stresses that probabilities and well-formed output do not guarantee a correct answer, and advises evaluating on the customer's own data with review of ambiguous or consequential cases. The service is proprietary and hosted only.