Constrained Decoding Models (CDMs) are ScaleDown's narrow 🏷️#ai-model family, each trained for a single task and restricted to returning only the answers that task allows.
ScaleDown offers four CDMs — Summarize, Compress, Extract, and Classify. Their output is limited to valid answers, such as a label from a supplied list or spans copied from the input, and each accepts up to 1M input tokens, including images and scanned PDFs. ScaleDown reports them as up to 200x faster and 467x cheaper than a 📝Frontier Model on narrow tasks; these are vendor benchmarks, not independent results.
CDMs differ from 📝Large Language Models (LLMs) in scope rather than kind of use. A general-purpose LLM writes any string one token at a time, which suits open-ended work like coding and planning. A CDM gives up that range for speed, cost, and outputs that cannot break schema. A 📝Langchain post from October 13, 2026 recommends pairing them, "cheap by default, frontier on exception": the LLM handles the hard reasoning while CDMs run the high-volume context steps of a long-running agent, such as summarizing old history and compressing large tool results.
CDMs sit in the same category as 📝Jev, TypeSafe AI's decision model. ScaleDown Classify most closely mirrors it, returning a typed answer with a probability. Classify also adds a one-sentence reason and accepts up to 1M tokens with images, where Jev takes text only, up to 32K tokens of state per question.