Daily Digest
CDM vs LLM vs Jev: picking the right model type
Decide whether a task needs an LLM, Jev, or a ScaleDown CDM.
Three kinds of model, one decision
A new label is showing up next to "LLM" in AI tooling: CDM. If you already read our System One / Jev deep dive, you have most of the background. This piece puts the three side by side so you can pick the right one for a job.
What each one is
LLM (large language model). A general model that writes its answer one token (word piece) at a time. It can write, explain, code and reason through open-ended questions. That flexibility is also its cost: every answer is generated text, which takes time and money, and any structure has to be parsed out afterwards (LangChain).
Jev (a "typed decision" model). Built by TypeSafe AI, which calls the category "System One" models. You send some text or JSON plus typed questions, and Jev returns answers of exactly that type, such as a choice from options you list (up to 255), a score on a rubric, or a yes/no probability. It does not write free text (TypeSafe docs; TypeSafe API reference).
CDM (Constrained Decoding Model). A brand name from ScaleDown, launched on 13 October 2026 (LangChain). Each model does one task (Classify, Extract, Summarize and Compress) and only returns answers that task allows, such as a label from your list or spans copied from the input (LangChain).
Three naming cautions:
- It is a vendor term, not an industry category. Even partners vary: LangChain's own launch post says "constrained decision models" (LangChain on LinkedIn), and one analysis notes ScaleDown's docs may apply the name to the Classify model specifically (smiit.in).
- It is not the general "constrained decoding" technique. Libraries like Outlines stop a normal LLM from writing tokens that break your format. The LLM still writes token by token. ScaleDown uses the phrase as a product name (smiit.in).
- Not every CDM is non-generative. Summarize writes a new summary (up to 1,000 tokens) rather than picking from the input (smiit.in).
What they output and where they fit
| LLM | Jev | CDM (ScaleDown) | |
|---|---|---|---|
| Output | Free text | Typed answer + probability (choice, score, yes/no) | Label + probability, input spans + confidence, or a short summary |
| Writes free text? | Yes | No | Only Summarize |
| Typical job | Writing, reasoning, code, agents | Routing, scoring, guardrails inside your code | Classify or extract over long documents; compress agent context |
| Input (vendor docs) | Varies by model | Up to 32K tokens per question; LangChain says text only, but TypeSafe's docs now also describe image input | Up to 1M tokens, images and scanned PDFs too |
| Where speed/cost numbers come from | Model provider | TypeSafe (company-reported) | ScaleDown (company-reported) |
Input rows are from LangChain's comparison and TypeSafe's docs.
Jev and CDMs overlap most on classification: LangChain says ScaleDown Classify "will feel familiar" to Jev users, since both return a typed answer with a probability, and Classify can add a one-sentence reason (LangChain). Neither replaces an LLM for open-ended work; both work best as routers, gates and validators around one (GPTS24).
The numbers, and whose they are
- ScaleDown (company-reported): 200x faster and 467x cheaper than traditional LLM APIs, 77.2% on DecisionBench vs Jev's 72%, and $0.05 per million input tokens with output free (launch post via LangChain on LinkedIn).
- TypeSafe (company-reported): large speed and cost multiples versus LLM workflows on its own evaluations; SiliconANGLE notes these were not independently verified (SiliconANGLE).
We found no independent head-to-head benchmark of CDMs vs Jev vs LLMs on the same tasks.
How to choose
- Does the answer need to be written? Emails, explanations, code, plans: use an LLM.
- Is the answer one of a known set (a label, a score, yes/no) and you want a probability to branch on? Try Jev, or ScaleDown Classify.
- Is the input long (contracts, PDFs, long agent history) and you need labels or exact fields pulled out? That is where CDM Classify/Extract is aimed, given its larger documented input (LangChain).
- Is the task new or fuzzy? Stay with an LLM. Task-specific models can fail abruptly on inputs unlike their training data, confidently misclassifying where an LLM might ramble (GPTS24).
Limits
- Vendor claims. Every speed, cost and accuracy number above comes from the company selling the model.
- New products. ScaleDown's CDMs launched on October 13, 2026 and Jev in mid-September 2026, so pricing, access and limits may change.
- No independent benchmark found. Treat comparisons, including DecisionBench, as directional.
- Typed isn't the same as correct. A model that can only return a valid label can still return the wrong one. Check it against your own labelled examples.
Try it
Take 50 past items you have already labelled by hand, such as support tickets with the right queue. Run them through your current LLM and through one typed option (Jev or ScaleDown Classify) using the same label list. Compare accuracy, cost and time on your data before moving any live traffic. Use non-sensitive samples for the first run.
Sources
- LangChain — Building Long-Running Agents with ScaleDown and Deep Agents (13 Oct 2026): https://www.langchain.com/blog/building-long-running-agents-with-scaledown-and-deep-agents
- LangChain on LinkedIn — partner post resharing ScaleDown launch: https://www.linkedin.com/posts/langchain_excited-to-partner-with-scaledown-their-activity-7515747815353159681-ugpy
- smiit.in — Pointing vs Generating: https://smiit.in/writing/answer-is-in-the-input/
- GPTS24 — ScaleDown's New AI Models Skip Token Generation: https://www.gpts24.com/en/news/scaledown-s-new-ai-models-skip-token-generation-to-cut-classification-costs
- Outlines (constrained decoding library): https://github.com/dottxt-ai/outlines
- TypeSafe docs — Introduction: https://docs.typesafe.ai/introduction
- TypeSafe docs — API reference: https://docs.typesafe.ai/llms-full.txt
- SiliconANGLE — TypeSafe exits stealth (16 Sep 2026): https://siliconangle.com/2026/09/16/typesafe-ai-exits-stealth-with-40m-to-build-ai-for-use-by-software/
- Daily Digest — System One / Jev deep dive: /daily-digest/system-one-and-jev-deep-dive