ScaleDown launches Constrained Decoding Models: 200x faster, 467x cheaper than frontier LLM APIs
omarsar0 · x · 2026-10-13
ScaleDown has released a family of three Constrained Decoding Models (CDMs) with 1M-token multimodal input for classification, extraction, and summarization.
- Core idea: LLMs are built to write, but most production workloads just need an answer. CDMs are purpose-built per task, read the input once, and point straight to the answer without token-by-token generation
- Claimed performance: 200x faster and 467x cheaper than frontier LLM APIs
- The three live models: Classify (calibrated probability + reason per label, for routing/tagging/intent), Extract (natural-language entity descriptions to well-formed JSON), and Summarize
More from Infra
- Decision model cuts main LLM calls 29.6% and cost 26.6% on 360 SOP-Bench tasks — vivekhaldar · 2026-10-13
- Budget 92GB VRAM rig: 4x used RX 7900 XT + RTX 3060 for ~$4,300 — Miserable-School-665 · 2026-10-13
- WSJ: the AI buildout is now the biggest investment boom in U.S. history — rand_longevity · 2026-10-13
- Modal publishes the manual for serving trillions of tokens to coding agents — charles_irl · 2026-10-13
- Basalt: A fast Blackwell inference engine for Qwen 3.8 Flash Next with dynamic vision — jesdga95 · 2026-10-13
- 80B MoE Model Runs in a Chrome Tab: llama.cpp WebGPU Backed Pages Experts From Disk — naklitechie · 2026-10-13