--- title: Jev skips token generation entirely, and the speed... description: TypeSafe AI launched Jev, a model that skips token generation and instead returns typed decisions with probabilities and confidence scores for structured... canonical: https://daily.dev/posts/jev-skips-token-generation-entirely-and-the-speed-numbers-are-hard-to-ignore-nv8tizgl7 twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Jev skips token generation entirely, and the speed numbers are hard to ignore | daily.dev og:description: TypeSafe AI launched Jev, a model that skips token generation and instead returns typed decisions with probabilities and confidence scores for structured... og:url: https://daily.dev/posts/jev-skips-token-generation-entirely-and-the-speed-numbers-are-hard-to-ignore-nv8tizgl7 og:image: https://api.daily.dev/og/posts/Nv8TIzgl7.png og:image:alt: Jev skips token generation entirely, and the speed numbers are hard to ignore 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 skips token generation entirely, and the speed numbers are hard to ignore **[Trends](https://daily.dev/sources/trends)** · 3 min read · 4 upvotes · 0 comments ## Summary TypeSafe AI launched Jev, a model that skips token generation and instead returns typed decisions with probabilities and confidence scores for structured questions, framed as a "smart if-statement" for workflows that need fuzzy judgment calls rather than prose. Claimed benchmarks: 70-500ms responses, 20-200x faster and 40-400x cheaper than comparable LLMs, trained via a calibration approach called RLCD. Early reactions from Santiago Valdarrama and Omar Syed praised the confidence-score routing and parallel structured-question design as architecturally novel, while cautioning that the speed and quality claims are self-reported and unverified pending independent benchmarks. ## Content TypeSafe AI launched Jev this week and the builder crowd is losing it — in a good way. The model doesn't generate text. It takes a structured question, evaluates it in parallel, and returns a typed decision with a confidence score in 70-500ms. That's it. No tokens streamed, no parsing, no hallucinated JSON. The numbers people are posting are hard to ignore. @nutlope classified 1,018 AI research papers for $0.08 total. He ran a fraud detection pipeline — 100 emails, 96% accuracy, 16 seconds end-to-end — for $0.07, routing low-confidence cases to Kimi K3 as a fallback. @iannuttall dropped it into Cloudflare Workers and got 7x faster search reranking and 50x faster content tagging versus GLM 4.7 Flash. Vercel's @rauchg says Jev is 18x faster (p95) *and* more accurate than GPT Luna for their safety reviewer, and it's likely becoming the new default in their AI Gateway. The pricing is what's making people do double-takes: $42 per billion input tokens, output effectively free. For comparison, that's roughly 40-400x cheaper than running the same classification through a general-purpose LLM. The skeptics have a point worth hearing though. One technical post argues Jev's speed advantage isn't a fundamentally new architecture — it's mostly single-token generation against constrained choices, something you can approximate by prefilling a regular LLM and restricting output. @scaling01 put it bluntly: "it's not a general language model and can't generate free form text... it can only output a few different primitives." @jerryjliu0 noted you can technically turn it back into an autoregressive LLM by letting it output ASCII, which is either a fun hack or a sign the moat is thinner than the launch post implies. The honest framing is probably this: Jev isn't replacing GPT-4 or Claude. It's replacing the part of your pipeline where you're burning $0.50 of inference to answer "is this ticket urgent?" The pattern people are converging on — fast specialized model for the narrow call, route uncertain cases to a bigger LLM — is genuinely useful and underexplored. Founded by Diogo Almeida, a former OpenAI researcher and RLHF co-inventor, TypeSafe is an early-stage startup, which means the usual caveats apply: single-region deployment, vendor dependency risk, and probability calibration that hasn't been stress-tested in regulated industries. The benchmarks are internal. Production reliability at scale is still an open question. But the use cases clicking into place fast — agent routing, LLM-as-judge evaluation, fraud detection, QA automation — suggest this fills a real gap. The question is whether the incumbents ship something equivalent before TypeSafe builds a moat. ## Questions this post answers ### What is Jev and how is it different from a typical LLM? Jev, launched by TypeSafe AI, does not generate text token by token. Instead it takes structured questions and returns typed decisions with probabilities and confidence scores, aimed at use cases like classifying support tickets as urgent or checking policy violations. Claimed performance is 70-500ms response times, 20-200x faster and 40-400x cheaper than comparable LLMs, since output tokens are free. _Developers weighing structured-decision models over text-generating LLMs can track how this plays out on daily.dev._ ### What training approach does Jev use to calibrate its confidence scores? Jev uses an approach TypeSafe AI calls RLCD, which calibrates probabilities to reflect how often the model is actually correct. This lets a confidence score double as a routing signal: when the model isn't sure, the output can be sent to human review instead of trusted directly, which Santiago Valdarrama highlighted as the interesting architectural piece. _Teams designing human-in-the-loop review pipelines can follow calibration approaches like this via daily.dev._ ## Community take How the wider developer community reacted, aggregated from 2 discussions and 6 comments across x (as of 2026-09-18). **TL;DR:** Replies mostly debate what layer of the stack Jev's speed gains actually belong to (architecture vs. routing vs. posttraining), with little direct evaluation of the product itself. **Sentiment:** 15% positive · 70% mixed · 15% skeptical **The case for** - Task-specific architectures make sense when outputs have a fixed schema and errors are measurable. **The pushback** - Some question whether Jev is really a new architecture versus just another harness or posttraining tweak. - Routing decisions at the harness layer can obscure which cost/latency/accuracy tradeoff is actually being chosen. **By community** - x (mixed): Replies focus on categorizing where Jev's approach fits (architecture, routing, or posttraining) rather than endorsing or rejecting its results, with one comparing it to a related local-model research effort. **Hottest debate:** Whether Jev represents a genuinely new architecture for classification or is just routing/harness-level optimization dressed up as one. **Open questions** - Is Jev's speed advantage due to a novel architecture or simply smart routing between models? **Highlights** > @jerryjliu0 so jev is a whole new architecture for classification, not just another harness or posttrain tweak > — [Flextor97 on x](https://x.com/Flextor97/status/2100786760984764549) > @jerryjliu0 Routing at the harness layer can hide the product decision, because choosing frontier or open-weight models per subtask quietly sets which accuracy, cost, and latency tradeoff wins. > — [mktpavlenko on x](https://x.com/mktpavlenko/status/2100792277140390089) > @jerryjliu0 This is very close to what we’ve been exploring with ProgramAsWeights: train a larger model to generate a small model’s LoRA adapter from an English description of the function. The resulting function runs locally on CPU and can be reused inside ordinary code. We’ve released the > — [yuntiandeng on x](https://x.com/yuntiandeng/status/2100803061304156378) > @jerryjliu0 Task-specific architectures fit best where outputs have a fixed schema and errors are measurable. > — [luyileo on x](https://x.com/luyileo/status/2100786722196144579) **Source threads** - [x](https://x.com/jerryjliu0/status/2100777020330491984) · 0 points · 6 comments - [x](https://x.com/jerryjliu0/status/2100819471795642462) · 0 points · 0 comments --- Tags: [#ai-agents](https://daily.dev/tags/ai-agents), [#ai-inference](https://daily.dev/tags/ai-inference) [View this post on daily.dev](https://daily.dev/posts/jev-skips-token-generation-entirely-and-the-speed-numbers-are-hard-to-ignore-nv8tizgl7) ```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 skips token generation entirely, and the speed numbers are hard to ignore","url":"https://daily.dev/posts/jev-skips-token-generation-entirely-and-the-speed-numbers-are-hard-to-ignore-nv8tizgl7","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/jev-skips-token-generation-entirely-and-the-speed-numbers-are-hard-to-ignore-nv8tizgl7"},"datePublished":"2026-09-15T19:40:24.942Z","dateModified":"2026-09-18T09:47:18.665Z","description":"TypeSafe AI launched Jev, a model that skips token generation and instead returns typed decisions with probabilities and confidence scores for structured...","isAccessibleForFree":true,"articleSection":"Trends","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":"Trends","logo":"https://media.daily.dev/image/upload/s--ZfSp3asX--/f_auto,q_auto/v1780996004/logos/trends?_a=BAMAMiWQ0","url":"https://daily.dev/sources/trends"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/jev-skips-token-generation-entirely-and-the-speed-numbers-are-hard-to-ignore-nv8tizgl7","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":4},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai-agents,ai-inference","timeRequired":"PT3M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Trends","item":"https://daily.dev/sources/trends"},{"@type":"ListItem","position":3,"name":"Jev skips token generation entirely, and the speed numbers are hard to ignore"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/jev-skips-token-generation-entirely-and-the-speed-numbers-are-hard-to-ignore-nv8tizgl7#faq","mainEntity":[{"@type":"Question","name":"What is Jev and how is it different from a typical LLM?","acceptedAnswer":{"@type":"Answer","text":"Jev, launched by TypeSafe AI, does not generate text token by token. Instead it takes structured questions and returns typed decisions with probabilities and confidence scores, aimed at use cases like classifying support tickets as urgent or checking policy violations. Claimed performance is 70-500ms response times, 20-200x faster and 40-400x cheaper than comparable LLMs, since output tokens are free. Developers weighing structured-decision models over text-generating LLMs can track how this plays out on daily.dev."}},{"@type":"Question","name":"What training approach does Jev use to calibrate its confidence scores?","acceptedAnswer":{"@type":"Answer","text":"Jev uses an approach TypeSafe AI calls RLCD, which calibrates probabilities to reflect how often the model is actually correct. This lets a confidence score double as a routing signal: when the model isn't sure, the output can be sent to human review instead of trusted directly, which Santiago Valdarrama highlighted as the interesting architectural piece. Teams designing human-in-the-loop review pipelines can follow calibration approaches like this via daily.dev."}}]} ```