Scaledown Launches CDMs, a Family of Constrained-Decoding Models That Skip Token-by-Token Generation

Scaledown team released the Constrained Decoding Models (CDMs) family on October 13, 2026, introducing three models — Classify, Extract, and Summarize — each purpose-built for a single task. Unlike conventional autoregressive LLMs that generate responses token by token, CDMs read the input once and skip the generative loop to achieve faster and more affordable inference.

Classify lets users define custom labels and returns calibrated probabilities with reasons, suitable for routing, tagging, and intent detection. Extract takes natural-language descriptions of desired entities and returns well-formed JSON with confidence scores and source citations. Summarize converts long text, images, or PDFs into summaries under 1,000 tokens, useful for context compaction and transcript digests.

The models support up to 1 million multimodal input tokens (text and image), score 77.2% on DecisionBench, and are claimed to be 200x faster and 467x cheaper than frontier LLM APIs. Input pricing is set at $0.05 per million tokens, with free output. The models are now live on platforms including OpenRouter, Vercel, LiteLLM, and LangChain.