ScaleDown Constrained Decoding Models

ScaleDown's small non-generative models for classification, extraction, and summarization that return only allowed outputs.

ScaleDown · Infrastructure and inference/serving · model

Published in Feed

This entry is marked for review. Check the original sources before relying on its details.

Overview

ScaleDown announced Constrained Decoding Models on October 13, 2026, a lineup of three small task-specific models (Classify, Extract, Summarize) that read the input once and return only an answer the task allows, such as a label from a supplied list or spans copied from the input. ScaleDown claims 200 times faster inference and 467 times lower cost than frontier LLMs on these narrow tasks and offered 50M free input tokens. LangChain published a launch post the same day and ships a langchain-scaledown package with summarization, tool-result compression, and extraction middleware; the models are also reachable through OpenRouter and Vercel. The speed and cost figures are vendor claims, and one critique notes constrained decoding is an existing technique and that the paid extraction model leads the free GLiNER2.5 by only 0.2 points on ScaleDown's own benchmark.

This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.

Key features

  • Three task models: Classify, Extract, and Summarize
  • Returns only task-allowed outputs (label from a list, or spans of the input) instead of free text
  • Vendor-claimed 200x faster and 467x cheaper than frontier LLMs on narrow tasks
  • langchain-scaledown middleware for summarization, tool-result compression, and extraction
  • Available via OpenRouter, Vercel, and LangChain; 50M free input tokens at launch

Use cases

  • High-volume document classification and field extraction
  • Compressing long tool outputs inside long-running agents
  • Cheap guardrail and routing steps in agent pipelines

Pricing

Managed API reported at $0.05 per 1M input tokens with output free (third-party report); 50M free input tokens at launch.

Pricing reflects the source record and may change. Confirm current terms with the provider.

Dates and sources

Released
13 Oct 2026
Record updated
18 Oct 2026
Record last checked
18 Oct 2026

Explore the Feed

Other entries filed under infrastructure and inference/serving; inclusion does not imply equivalent features.

  • NVIDIA PAIR (Personal AI Router)

    NVIDIA's free, open-source personal AI router that spreads local model workloads across PCs on your network; out of beta Oct 13, 2026.

  • Agent37

    YC F26 startup selling low-cost agent sandboxes (gVisor or full VM) with prebuilt harnesses for Claude Code, Hermes and opencode.

  • Microsoft-Decision-1

    Microsoft's Qwen3.5-9B-based decision-scoring model returns calibrated option probabilities in one pass, at $0.042 per million input tokens.

Browse the Feed