Scaledown is launching Constrained Decoding Models (CDMs) , a model architecture aimed at production AI workloads where the application already knows the shape of the answer it needs. The initial release includes three task-specific small language models: Classify, which routes text by assigning user-provided labels along with a probability for each label, Extract, which reads files and outputs the requested data as JSON fields with a confidence score, and Summarize, which condenses long transcripts or reports into summaries of up to 1,000 tokens . The pitch to developers is that classification, extraction, and summarization do not require generating a full LLM response token by token when the output structure is already constrained.
The models are built for high-volume backend jobs rather than open-ended chat. All three support a 1M-token input limit and accept text or images, and they are accessible through ScaleDown's API as well as OpenRouter, Vercel AI Gateway, LiteLLM, and LangChain . Pricing is unusual for the category: input is billed at $0.05 per million tokens, with a promotional $0.04 on OpenRouter at launch, and output tokens are not counted . On accuracy, ScaleDown reports 95% on the CLINC OOS public intent-detection benchmark for Classify, versus an average of 89.6% for three OpenAI models.
Scaledown is a New York-based applied AI research lab founded in 2024, building task-specific small language models for compression, summarization, extraction, and classification , and is led by co-founder and CEO Neal Patel, a UC Berkeley alum. The launch matters now because teams running classification pipelines, document extraction, and ticket routing at scale are actively looking for cheaper alternatives to frontier models on jobs where a general-purpose chat model is more capability than the task requires.
Comments (11)
Sign in to join the discussion.
Y
Yuki Tanabe20h ago
If you're actually cutting token spend on classification workloads, every ops team with a RAG pipeline should be forwarding this to their CFO by Monday. Curious what the pricing floor looks like before margin collapses.
C
Chidera Obiagu20h ago
The cold open on that launch video is doing a lot of work. Zero intro slide, straight into the problem statement, more founders should steal this pacing.
R
Rita Vaz19h ago
Neal buried the lede by waiting until sentence three to say 'new architecture'. That belongs in the first six words of the hook.
P
Parv Rajagopal19h ago
dumb intern question but if the model already knows the answer schema, isn't this just... really fancy regex with embeddings? what am i missing
L
Lena Bergström19h ago
Reminds me of what Modal and Baseten were pitching early on, infra-level wins that sound boring until your bill shows up. Watching this one.
K
Koa Nakamura19h ago
Hit the docs before the tweet finished loading. Quickstart is clean but I want a self-host story and a Discord link before I evangelize this internally.
J
Jax Hwang19h ago
webhook support? streaming? what happens if my enum has 400 values and i'm doing it under 50ms p99? asking for a very specific friend
S
softboot19h ago
Every LLM pretending to think when it's really just picking from four options. The emperor finally gets a fitted suit.
A
Amara Okonkwo19h ago
Poked at the playground on mobile and the first run errored silently with no toast. Fix that before the HN crowd shows up tomorrow morning.
M
Mikhail Varga19h ago
been building something adjacent for ~3 weeks, was gonna launch next month but now i need to rethink my entire positioning deck lol. gg
N
Nadia El-Sayed19h ago
Tweet is sitting at a weirdly high impression count relative to replies, which usually means people are reading, bouncing, then DMing. That's a strong signal for infra launches actually.
Work With Us
Book an intro call to explore our suite of services. We might have a waitlist.
Scaladown
Constrained Decoding Models
Constrained Decoding Models
Tags
TLVC Rating
Community Rating
No ratings yet
to rate this launch.
Scaledown is launching Constrained Decoding Models (CDMs) , a model architecture aimed at production AI workloads where the application already knows the shape of the answer it needs. The initial release includes three task-specific small language models: Classify, which routes text by assigning user-provided labels along with a probability for each label, Extract, which reads files and outputs the requested data as JSON fields with a confidence score, and Summarize, which condenses long transcripts or reports into summaries of up to 1,000 tokens . The pitch to developers is that classification, extraction, and summarization do not require generating a full LLM response token by token when the output structure is already constrained. The models are built for high-volume backend jobs rather than open-ended chat. All three support a 1M-token input limit and accept text or images, and they are accessible through ScaleDown's API as well as OpenRouter, Vercel AI Gateway, LiteLLM, and LangChain . Pricing is unusual for the category: input is billed at $0.05 per million tokens, with a promotional $0.04 on OpenRouter at launch, and output tokens are not counted . On accuracy, ScaleDown reports 95% on the CLINC OOS public intent-detection benchmark for Classify, versus an average of 89.6% for three OpenAI models. Scaledown is a New York-based applied AI research lab founded in 2024, building task-specific small language models for compression, summarization, extraction, and classification , and is led by co-founder and CEO Neal Patel, a UC Berkeley alum. The launch matters now because teams running classification pipelines, document extraction, and ticket routing at scale are actively looking for cheaper alternatives to frontier models on jobs where a general-purpose chat model is more capability than the task requires.
Comments (11)
If you're actually cutting token spend on classification workloads, every ops team with a RAG pipeline should be forwarding this to their CFO by Monday. Curious what the pricing floor looks like before margin collapses.
The cold open on that launch video is doing a lot of work. Zero intro slide, straight into the problem statement, more founders should steal this pacing.
Neal buried the lede by waiting until sentence three to say 'new architecture'. That belongs in the first six words of the hook.
dumb intern question but if the model already knows the answer schema, isn't this just... really fancy regex with embeddings? what am i missing
Reminds me of what Modal and Baseten were pitching early on, infra-level wins that sound boring until your bill shows up. Watching this one.
Hit the docs before the tweet finished loading. Quickstart is clean but I want a self-host story and a Discord link before I evangelize this internally.
webhook support? streaming? what happens if my enum has 400 values and i'm doing it under 50ms p99? asking for a very specific friend
Every LLM pretending to think when it's really just picking from four options. The emperor finally gets a fitted suit.
Poked at the playground on mobile and the first run errored silently with no toast. Fix that before the HN crowd shows up tomorrow morning.
been building something adjacent for ~3 weeks, was gonna launch next month but now i need to rethink my entire positioning deck lol. gg
Tweet is sitting at a weirdly high impression count relative to replies, which usually means people are reading, bouncing, then DMing. That's a strong signal for infra launches actually.