Releases·Model Releasefeatured
TypeSafe AI launches Jev, an RLCD-trained decision model claiming 20-200x faster, 40-400x cheaper inference than LLMs
TypeSafe AI, founded by ex-OpenAI researcher Diogo Almeida, released Jev on September 15, 2026 after two years in stealth — a frontier model that takes unstructured state plus typed questions and returns parallel, calibrated decisions instead of generated text. It is trained with a new objective, Reinforcement Learning for Calibrated Decisions (RLCD), positioned against RLHF and RLVR. TypeSafe claims roughly two orders of magnitude faster and cheaper inference than LLMs on decision tasks, and community reaction has been intense: follow-up analyses, a JevBench benchmark, and open-weight imitators like Kev appeared within days.
1749 engagement17 sourcesTue, Sep 15, 2026, 12:00 AM
Developments
TechCrunch reports Vercel saw 5-18x speed gains replacing ChatGPT Luna 5.6 with Jev; TypeSafe's API briefly failed under launch demand.
Sat, Sep 19, 2026, 04:03 AM
OpenJev, an open-source server replicating Jev's API on DiffusionGemma via vLLM PR #57250, launched with near-parity eval accuracy.
Fri, Sep 18, 2026, 05:26 PM
AI Overview
- TypeSafe AI founder Diogo Almeida, an OpenAI RLHF co-inventor, left two years ago to build Jev, arguing optimizing for human language blocks automation
- Jev lets large LLMs handle hard reasoning while cheap models process millions of small decisions; Vercel replaced ChatGPT Luna 5.6 with Jev, getting results 5-18x faster with greater accuracy
- TypeSafe claims Jev is 20-200x faster (70-500ms) and 40-400x cheaper than LLMs, with workflow evals vs the average of GPT-6 Astra and Fable 5.1 backing homepage figures of 193.6x faster and 444.6x cheaper; input tokens cost $0.042/million, output tokens free
- Jev is trained via RLCD, cannot produce free-form text or explain its decisions, publishes no calibration measurements, and its docs state P(x) and P(not x) need not sum to 1; Willison flags hidden bias as a key risk
- An open source MCP/ACP bridge lets agents integrate Jev without building a custom adaptor
- Matt Mastracci's unmerged vLLM PR #57250 runs DiffusionGemma as a Jev-style one-step decision model, spawning OpenJev (codiv.ai, Apache-2.0, 100M tokens/account) and Kev, Qwen 3.5-based 0.8B/4B/9B clones; JevBench already benchmarks Jev-class models
- Mastracci's evals show DiffusionGemma-as-Jev roughly tied with Jev on accuracy (198/201 vs 191/201) and ran faster on a DGX Spark
- OpenJev serves ~170 ms p50 end-to-end latency on an RTX PRO 6000 (~73 ms on-GPU) vs Jev's published 70–500 ms
Entities
Related
Products·launchedfeatured
TypeSafe Jev: AI model that returns typed decisions with probabilities instead of text
Jev is a decision-only AI model from TypeSafe: you provide app state plus typed questions (yes/no, pick-one, score) and it returns answers with probabilities, no text to parse. It targets the many LLM calls in apps that are really just classification or routing decisions, offering a cheaper, faster alternative for developers building agentic or decision-driven software.
Jev is a decision-only AI model from TypeSafe: you provide app state plus typed questions (yes/no, pick-one, score) and it returns answers with probabilities, no text to parse. It targets the many LLM calls in apps that are really just classification or routing decisions, offering a cheaper, faster alternative for developers building agentic or decision-driven software.
34031 engagement8 sourcesreddit, rssWed, Sep 16, 2026, 11:42 PM
Releases·Model Release
Jev: System One models for Prod, not God — with Diogo Almeida, CEO, TypeSafe AI
During Jev launch week, Latent Space published an interview with Diogo Almeida, CEO of TypeSafe AI and InstructGPT co-author, on why he built a 'System One' model for production software rather than chat. Jev's launch video drew ~40M views, and its core training technique, Reinforcement Learning for Calibrated Decisions (RLCD), optimizes for calibrated probabilities instead of human preference feedback. Practitioners get official patterns, cookbooks, and a coding-agents guide, plus use cases spanning computer use, games, linting, and analytics.
During Jev launch week, Latent Space published an interview with Diogo Almeida, CEO of TypeSafe AI and InstructGPT co-author, on why he built a 'System One' model for production software rather than chat. Jev's launch video drew ~40M views, and its core training technique, Reinforcement Learning for Calibrated Decisions (RLCD), optimizes for calibrated probabilities instead of human preference feedback. Practitioners get official patterns, cookbooks, and a coding-agents guide, plus use cases spanning computer use, games, linting, and analytics.
0 engagement1 sourcerssMon, Sep 21, 2026, 10:13 PM
Social·Community
TypeSafe AI Launches Jev Decision Model, Drawing Developer Interest
Medium writers and Reddit users are discussing Jev, a decision model from TypeSafe AI introduced in September 2026. Writers describe it as answering with probabilities rather than words, choosing among provided options for small agent decisions, and potentially making decision intelligence practical via a cheaper, faster enterprise stack. Reddit users report running local Jev-like models for deep eval testing and ask peers for practical use cases, with one calling it a 'solution in search of a problem'.
Medium writers and Reddit users are discussing Jev, a decision model from TypeSafe AI introduced in September 2026. Writers describe it as answering with probabilities rather than words, choosing among provided options for small agent decisions, and potentially making decision intelligence practical via a cheaper, faster enterprise stack. Reddit users report running local Jev-like models for deep eval testing and ask peers for practical use cases, with one calling it a 'solution in search of a problem'.
060205 engagement6 sourcesrss, redditSat, Oct 10, 2026, 01:26 AM
Releases·Model Release
Amazon releases its own Jev clone as decision models flood the web
AWS released Strands Decider 2B, an open source decision model that sorts between pre-decided options with calibrated confidence scores instead of generating text, built on the Qen3.5-2B LLM torso. Distinguished engineer Marc Brooker built it after seeing TypeSafe's Jev, and it briefly topped the Jevbench ranking for its size class before Amazon released it via Strands Labs. The release lands the same week OpenAI announced a similar offering, amid a wave of decision-model clones following Jev's debut.
AWS released Strands Decider 2B, an open source decision model that sorts between pre-decided options with calibrated confidence scores instead of generating text, built on the Qen3.5-2B LLM torso. Distinguished engineer Marc Brooker built it after seeing TypeSafe's Jev, and it briefly topped the Jevbench ranking for its size class before Amazon released it via Strands Labs. The release lands the same week OpenAI announced a similar offering, amid a wave of decision-model clones following Jev's debut.
101 engagement2 sourcesrss, githubThu, Oct 1, 2026, 04:49 PM
Products·launched
OpenJev: Qwen3.5 4B logit-probability scoring replicating Jev-style choice models
OpenJev is an open implementation showing that Qwen3.5 4B with per-option logit probabilities can match or beat TypeSafe AI's Jev model, which outputs probabilities over choices. It offers a browser demo for trying it directly. It targets local LLM users wanting cheap multiple-choice probability scoring without a dedicated reranker.
OpenJev is an open implementation showing that Qwen3.5 4B with per-option logit probabilities can match or beat TypeSafe AI's Jev model, which outputs probabilities over choices. It offers a browser demo for trying it directly. It targets local LLM users wanting cheap multiple-choice probability scoring without a dedicated reranker.
68 engagement1 sourceredditWed, Sep 16, 2026, 03:10 PM