ScaleDown Unveils 3 Constrained Decoding Models: 1M Multimodal Context and Free Output Tokens

On October 13, 2026, artificial intelligence startup ScaleDown officially launched Constrained Decoding Models (CDMs), a novel model architecture purpose-built to eliminate the sequential token-by-token generative loop of standard large language models (LLMs). Tailored for deterministic production workloads, the initial lineup comprises three specialized models—Classify, Extract, and Summarize—supporting a 1-million (1M) token multimodal (text + image) context window, alongside PDF support in Summarize.

Image source: Neal Patel (@neal_k_patel) / ScaleDown

Standard frontier LLMs are fundamentally engineered for open-ended text composition. Consequently, even when the exact target answer is already contained within an input prompt or referenced file, the model still generates its response token by token in an auto-regressive loop. ScaleDown directly tackles this generation latency and compute overhead through single-pass constrained decoding, allowing the model to ingest the input once and output structured results without entering an iterative decoding cycle.

Bypassing the Generative Loop: Single-Pass Constrained Decoding Architecture

The core premise behind ScaleDown's Constrained Decoding Models lies in purpose-built task specialization.

While conventional LLMs are designed to write open-ended text, the overwhelming majority of production enterprise workloads require rapid, consistent, and cost-effective answers to discrete queries. CDMs are architected to read the full input context in a single pass, completely bypassing the token-by-token generative loop that drives up operational latency and cost in standard models.

According to internal benchmark data shared by ScaleDown, this architectural shift yields notable operational efficiency gains compared to commercial frontier LLM APIs:

• Inference Latency: Delivers 200x faster inference speeds compared to standard frontier LLM APIs
• Operational Cost: Delivers 467x lower operational costs compared to standard frontier LLM APIs
• Decision Accuracy: Achieves 77.2% on the DecisionBench benchmark, outperforming dedicated decision model Jev (72%)

Three Purpose-Built CDMs: Classify, Extract, and Summarize

Available immediately at launch, the three models in the initial CDM family each handle a dedicated structured execution pipeline.

1. Classify (Routing and Tagging)

Given a user-defined list of classification labels, Classify outputs calibrated probability scores alongside an explicit reasoning rationale for each label. Geared toward model routing, document tagging, and customer intent detection, the model supports single-pass multi-label classification across more than 20 distinct tags per input, moving beyond simple single-choice selection.

2. Extract (Structured Information Extraction)

Developers define target entities in natural language, and Extract returns well-formed JSON paired with confidence scores and source citations, designed for web scraping, document processing, and automated PII redaction workflows.

3. Summarize (Multimodal Document Summarization)

Summarize processes long-form text, images, and PDFs to deliver structured, consistent summaries constrained under 1,000 tokens. This makes it suitable for context compaction in multi-agent architectures, document metadata indexing, and meeting transcript digests.

1M Multimodal Context Window and $0.05/1M Input Pricing

The ScaleDown CDM suite pairs high-capacity processing specifications with a competitive pricing structure tailored for high-volume enterprise ingestion.

• 1M Token Multimodal Context Window: Ingests multimodal (text + image) input up to 1 million tokens in a single call. This represents a 31x capacity increase over Jev.
• Input-Only Billing and Free Output: Pricing is set at $0.05 per 1 million input tokens ($0.05/1M), while all generated output tokens—including those from Summarize—are provided entirely free of charge.
• Launch Promotion: ScaleDown is offering 50 million (50M) free input tokens during the first 50 hours following the official release.
• Ecosystem Integration: The models are live from day one on major AI routing and development platforms, including OpenRouter, Vercel, LiteLLM, LangChain, TrustedRouter, Arthur AI, and useRouterPlus.

To minimize prompt engineering friction, ScaleDown also published 'SLM Agent', an open-source GitHub utility that automates system prompt generation for custom label and entity definitions.

Production Boundaries and Benchmark Verification Realities

While the architectural efficiency of single-pass constrained decoding is compelling, production deployments carry distinct operational boundaries that engineering teams must evaluate.

First, CDMs are strictly single-purpose models optimized for classification, structured extraction, and summarization. They cannot function as general-purpose conversational chatbots, nor can they handle multi-turn dialog or open-ended creative composition.

Second, the performance figures cited by ScaleDown—including 200x speedups, 467x cost reductions, and a 77.2% score on DecisionBench—originate from internal vendor benchmarks. Independent third-party validation across varied real-world production datasets will be necessary to confirm consistency across edge cases.

Finally, when dealing with ambiguous source documents where answers are not explicitly documented, developers must define fallback guidelines directly within their system prompt definitions. Furthermore, while the Classify model can be deployed as an intent router today, ScaleDown indicated that dedicated router models and specialized tool-calling models remain in active development for future release.

Sources

• Neal Patel Official Announcement on X (@neal_k_patel): ScaleDown CDM Release Announcement
• ScaleDown Official Website: ScaleDown AI
• ScaleDown Team GitHub Repository: scaledown-team/SLM_Agent