--- title: Jev is an encoder-only AI model built for... description: Jev, a new model from Typesafe AI, is an encoder-only model built specifically for classification tasks rather than text generation. Instead of producing text... canonical: https://daily.dev/posts/jev-is-an-encoder-only-ai-model-built-for-classification-not-text-generation-uvnjpygpg twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Jev is an encoder-only AI model built for classification, not text generation | daily.dev og:description: Jev, a new model from Typesafe AI, is an encoder-only model built specifically for classification tasks rather than text generation. Instead of producing text... og:url: https://daily.dev/posts/jev-is-an-encoder-only-ai-model-built-for-classification-not-text-generation-uvnjpygpg og:image: https://api.daily.dev/og/posts/uvnJPYGPG.png og:image:alt: Jev is an encoder-only AI model built for classification, not text generation og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Jev is an encoder-only AI model built for classification, not text generation **[Collections](https://daily.dev/sources/collections)** · 7 min read · 1 upvotes · 0 comments ## Summary Jev, a new model from Typesafe AI, is an encoder-only model built specifically for classification tasks rather than text generation. Instead of producing text token by token, it outputs a structured decision - a choice, score, or boolean - in a single forward pass, making it fast and cheap (about 4 cents per million input tokens, free output tokens). It lacks chain-of-thought reasoning, so it won't replace general-purpose LLMs for reasoning tasks, but it's well-suited for narrow classification jobs like malicious URL detection or sentiment scoring. New output classes can be added without retraining the whole model. ## Content Jev is a model from TypeSafe AI that doesn't write text. You give it some context and a set of typed questions (a choice, a yes/no, a score), and it returns answers with calibrated probabilities. It launched in mid-September 2026. TypeSafe is led by former OpenAI researcher Diogo Almeida, and the model is billed as a "System One" model, after Kahneman's fast, automatic mode of thinking. ## How it works All answers come back in one parallel pass instead of token by token. That is why it is fast and cheap. The listed numbers are a 32K token context window, $0.042 per million input tokens, free output, and 70-500ms latency. One explainer says it is trained with RLCD, reinforcement learning for calibrated decisions. The goal is probabilities that reflect real confidence, rather than the confident-sounding wrongness RLHF-trained LLMs can produce. TypeSafe hasn't disclosed the method. One analysis relates it to the published RLCR technique, which uses calibration rewards. The structure also helps. Jev can't emit an invalid type, though it can still return a confidently wrong valid value. ## Adoption Vercel added Jev to AI Gateway on day two and reported nearly 13% of paid teams using it within 24 hours, twice the rate of the GPT-5.6 family. Netlify and LangChain integrated it too. Vercel's Python AI SDK exposes it through an experimental evaluate() API. @rauchg called a writeup on Jev with Python wonderful. Open and local options appeared within days: - Ollama 0.35 added a /v1/systemone endpoint. Three models ship at launch: Bespoke Labs' 9B nimble and Together AI's 4B and 0.8B tev1 models. Nimble averaged 91ms per decision on an M5 Max in a Pac-Man demo. - @mitsuhiko noted that llama.cpp with Pi lets you use any model as a discount Jev for classification. - Featherless released Simple Jev, an open-source library that turns open models into zero-shot classifiers, with vision inputs on its hosted endpoints via Gemma or Qwen. It starts at $0.03 per million input tokens. CEO Eugene Cheah compares using frontier LLMs for simple classification to using a tank to deliver a pizza. Its public demo endpoints are capped at 2,000 tokens of context and two requests per second. - DuckDB community extensions such as duckdb-jev and duck-jev add functions like jev(), jev_prob(), jev_choice() and jev_score() for plain-English conditions in SQL. - A German-language analysis lists Cloudflare's Clef, AWS's Strands Decider and OpenAI's Decisions API as alternatives that appeared within two weeks. ## What people are building **Agent harnesses.** Most steps in an agent loop are decisions, not text generation: tool selection, safety checks, ranking, triage. One walkthrough covers model routing, risk gating, skill selection with progressive disclosure, RAG re-ranking, ticket triage and real-time filtering. A LangGraph course builds a production harness around Jev for routing, risk checks and evidence verification. @svpino says he moved all his workflows to Jev and that they are faster and cheaper. Because every prediction carries a confidence score, he can branch on it. **RAG.** Jev can sit between retrieval and generation. After hybrid search produces candidates, it scores each passage in one batched request. Application code then applies thresholds to decide which passages reach the LLM and whether generation runs at all. That keeps the relevance policy visible and tunable in code instead of buried in a prompt. @marktenenholtz thinks one of its best uses right now is as a search reranker for tasks you won't run repeatedly in production. It also works in production if you don't want to fine-tune your own model and can accept lower performance. @starbuxman pointed to a post by Spring AI lead Christian Tsolov on modular RAG with Jev. **Product features.** - @TejasKumar_ uses Jev to block questions like "what is your salary at ibm" on his site's ask boxes. It stops them in about 260ms for $0.0002, where the request previously reached gpt-4o. - There There uses Jev, through the Laravel AI SDK, for smart workflow conditions written in plain language and for ticket assignment. Assignment uses a Choice question with a "none of these" fallback so vague tickets aren't misrouted. - A software agency built a résumé screener on Next.js, Supabase and Jev. The argument is that numbers generated by an LLM aren't real probabilities and can't be meaningfully sorted. The composite score is computed in application code, and low-confidence cases go to human review. **Analytics and documents.** MotherDuck's prompt_jev() classified 100,000 CFPB complaint narratives in 82.2 seconds for about $0.50. LLMs like gpt-5-nano and GPT-5.6 took 6+ minutes on far fewer rows, at $1.58-$37.58. It is available on paid plans and sends data to TypeSafe. @jerryjliu0 reports that Jev tops most benchmarks for document tasks like orientation detection, language detection, classification and splitting, across accuracy, cost and latency, against other open-source models. **LLM judging.** In a benchmark against GPT-OSS-120B on 72 harder QA answers in English and Japanese, Jev matched human labels on 64 and GPT-OSS on all 72. Jev was about seven times faster (0.20s vs 1.44s median) and cost roughly $0.020 per 1,000 judgments. Its misses were near-correct answers with one material error. Sending uncertain judgments (probability 0.2-0.8) to GPT-OSS caught every error and escalated fewer than 20% of cases. MLflow 3.17 will support Jev through a typesafe:/ model URI in make_judge. ## Limits and skepticism - Speed claims are contested. Developers report 5-18x speedups over comparable LLM classifiers. Independent analysis found a median closer to 7x, against a 193.6x headline figure. - A German-language analysis argues the hype, nearly 2,000 Hacker News points, comes from the interface rather than novelty. Zero-shot classification, structured outputs and calibration have existed for years. Its independent benchmarks (BANKING77, phishing datasets, logic tasks) show speed and cost claims holding only under specific conditions, with accuracy sometimes equal to or worse than GPT and Claude. The lasting contribution, it concludes, is the pattern of context plus typed questions plus probabilities. - Red Hat's AI Safety team tested Jev on prompt-injection and content-safety guardrail benchmarks. Pre-trained CPU-scale classifiers stayed competitive on accuracy and latency without GPUs. Jev didn't reliably beat LLM-as-a-judge or open decision models. Prompt engineering mattered a lot: one tuned policy gained nearly 18 points. Their conclusion is that decision models are a good option when no pre-trained classifier exists, and that you should use the right tool for the job. - Documented weak spots include counting, arithmetic, date comparisons, large context, compound questions, text-only input (no images in Jev itself) and prompt injection. - DuckDB users note that most benchmarks come from extension authors, accuracy drops at larger batch sizes, and fine-tuned ModernBERT or embedding retrieval may beat Jev for fixed, high-volume classification. ## Where it fits Jev has no reasoning step. It produces its answer in one shot, so partial outputs can't be evaluated or improved mid-generation. That is why it won't replace GPT for reasoning-heavy work. The pattern that keeps coming up is Jev for fast triage and LLMs for the slower thinking. One piece frames it as a middle step between raw LLM prompting and a custom classifier: start with an LLM, collect real data, write a rubric, iterate, and train your own classifier once the problem is understood. On the research side, @lateinteraction shared a thread asking whether Jev is secretly learning a value function, since a model can choose "A" without being certain "A" is correct. I find that question more interesting than the speed numbers. ## Questions this post answers ### What is an encoder-only AI model and how is it different from GPT-style models for classification tasks? An encoder-only model like Jev generates a structured decision - a choice, score, or boolean - in a single forward pass rather than decoding text token by token. This makes it faster and cheaper than general-purpose LLMs for tasks like classification, since there is no sequential decoding loop. It lacks chain-of-thought reasoning, so it cannot handle tasks requiring step-by-step reasoning. _daily.dev surfaces architecture tradeoffs like this for developers choosing between LLMs and classifiers._ ### How much does the Jev classification model cost to run? Input tokens cost about 4 cents per million, and output tokens are free, since the model returns a single structured value like a score or boolean rather than a stream of generated text. This pricing structure reflects its encoder-only design, which produces one decision per forward pass instead of decoding text sequentially. _Developers comparing model costs for classification workloads can track pricing shifts like this on daily.dev._ ## Community take How the wider developer community reacted, aggregated from 2 discussions and 2 comments across x (as of 2026-10-05). **TL;DR:** There's very little independent discussion to go on; the few replies found focus narrowly on speed benefits and how the modular retrieval/reranking/generation split fits into existing pipelines. **Sentiment:** 40% positive · 60% mixed · 0% skeptical **The case for** - Speed advantages of these encoder-only models over traditional LLMs were highlighted. - The modular split (retrieval, reranking, generation as separate swappable stages) was seen as a good fit for existing app architectures, easing stage-by-stage testing. **By community** - x (mixed): Minimal engagement, with brief notes on speed benefits and architectural fit rather than any real debate. **Highlights** > @starbuxman The modular split is a strong fit for Spring apps: keep retrieval, reranking, and generation as swappable stages, then test each stage with traceable IDs. That makes it much easier to tune recall without hiding failures behind one giant chain. > — [gagu\_sh on x](https://x.com/gagu_sh/status/2106086559384842351) > @starbuxman We actually went deeper on the speed advantages of these new models here: > — [MichaelWaitze on x · 1 points](https://x.com/MichaelWaitze/status/2106082217353150820) **Source threads** - [x](https://x.com/rauchg/status/2106139305131569178) · 0 points · 0 comments - [x](https://x.com/starbuxman/status/2106050776032256099) · 0 points · 2 comments --- Tags: [#machine-learning](https://daily.dev/tags/machine-learning), [#ai-inference](https://daily.dev/tags/ai-inference), [#jev](https://daily.dev/tags/jev) [View this post on daily.dev](https://daily.dev/posts/jev-is-an-encoder-only-ai-model-built-for-classification-not-text-generation-uvnjpygpg) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Jev is an encoder-only AI model built for classification, not text generation","url":"https://daily.dev/posts/jev-is-an-encoder-only-ai-model-built-for-classification-not-text-generation-uvnjpygpg","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/jev-is-an-encoder-only-ai-model-built-for-classification-not-text-generation-uvnjpygpg"},"datePublished":"2026-09-28T19:09:19.221Z","dateModified":"2026-10-05T11:27:23.657Z","description":"Jev, a new model from Typesafe AI, is an encoder-only model built specifically for classification tasks rather than text generation. Instead of producing text...","image":"https://i.ytimg.com/vi/d9INaNO3Nrs/sddefault.jpg","thumbnailUrl":"https://i.ytimg.com/vi/d9INaNO3Nrs/sddefault.jpg","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/jev-is-an-encoder-only-ai-model-built-for-classification-not-text-generation-uvnjpygpg","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"machine-learning,ai-inference,jev","timeRequired":"PT7M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"Jev is an encoder-only AI model built for classification, not text generation"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/jev-is-an-encoder-only-ai-model-built-for-classification-not-text-generation-uvnjpygpg#faq","mainEntity":[{"@type":"Question","name":"What is an encoder-only AI model and how is it different from GPT-style models for classification tasks?","acceptedAnswer":{"@type":"Answer","text":"An encoder-only model like Jev generates a structured decision - a choice, score, or boolean - in a single forward pass rather than decoding text token by token. This makes it faster and cheaper than general-purpose LLMs for tasks like classification, since there is no sequential decoding loop. It lacks chain-of-thought reasoning, so it cannot handle tasks requiring step-by-step reasoning. daily.dev surfaces architecture tradeoffs like this for developers choosing between LLMs and classifiers."}},{"@type":"Question","name":"How much does the Jev classification model cost to run?","acceptedAnswer":{"@type":"Answer","text":"Input tokens cost about 4 cents per million, and output tokens are free, since the model returns a single structured value like a score or boolean rather than a stream of generated text. This pricing structure reflects its encoder-only design, which produces one decision per forward pass instead of decoding text sequentially. Developers comparing model costs for classification workloads can track pricing shifts like this on daily.dev."}}]} ```