TypeSafe AI releases Jev, a ‘System One’ decision model that returns typed probabilities instead of text
Confirmed
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- 3 of 5
- Last verified
- 17 October 2026
Your AI and this story
- GPT-6.2 Astra107 days after its cutoff
- Claude Opus 5.577 days after its cutoff
- Gemini 4 Hydrogen168 days after its cutoff
- Grok 4.7107 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 77 days before it.
Key facts
- Oct 9, 2026 (Bloomberg): TypeSafe raised ~$870M at a $7.5B valuation led by Andreessen Horowitz, with Sequoia and angels; company says Jev passed 1M users within days and is used by about a third of the Fortune 500 (see TypeSafe AI, maker of the Jev decision model, raises ~$870M led by a16z at a $7.5B valuation)
- Oct 2 (WSJ): CEO Diogo Almeida says Jev is in use by about 25% of Fortune 500 companies and ‘we were at a trillion tokens per day about a week ago’, traffic driven by servers rather than people; company figures, not independently verified, and ‘in use’ is undefined
- Released Sept 15, 2026 (TypeSafe blog), early access at launch
- Question types: Boolean (probability 0–1), Choice (distribution over options), Score (numeric rating)
- Latency 70–500 ms; TypeSafe claims up to 193.6x faster and 444.6x cheaper than LLMs on its workflow evals, with intelligence similar to GPT-5.6 Terra on ‘System One tasks’
- Pricing: $0.042 per 1M input tokens; output free
- TypeSafe: ‘unstructured state in, typed probabilistic decisions out’; claims no hallucinated or mistyped outputs
- Available through TypeSafe’s API (jev-latest), Vercel AI Gateway (typesafe-ai/jev) and Cloudflare (typesafe/jev)
Show 16 more
- Simon Willison proposed the name ‘decision models’ and warned the numbers ‘could conceal all manner of unseen bias’
- Reported $40M seed led by DCVC (not confirmed from TypeSafe’s own post)
- Founder Diogo Almeida’s launch post on X drew ~40M views by Sept 29; Vercel said Jev reached ~13% of AI Gateway teams on day one, ‘2x the GPT-5.6 family and 6x Fable 5.1’, its fastest adoption ever
- Fast followers: Cua open-sourced CUA-S1-FORMS, a 706K-parameter ‘System One’ model for form filling (99.7% vs hosted Jev’s 83.6% on Cua’s own eval); Bespoke Labs’ open Nimble was pitched as an alternative
- AWS follower (Oct 1, 2026): Strands Labs, AWS’s experimental agent project, released Strands Decider 2B (Apache 2.0), a Qwen3.5-2B-Base LoRA with a pointer/readout head that scores supplied options instead of generating text. Model card: JevBench public (231 tasks) 0.723, ContractNLI 0.872, MuSiQue 0.884. VentureBeat: 106 ms median and 296 ms p95 latency on an RTX 3090; #2 among public ~2B models on JevBench and #1 with a full training recipe
- Cloudflare follower (Oct 1, 2026): Clef (Qwen3.8-27B) and Clef-flash (Qwen3.5-9B), Apache-2.0 multimodal decision models, fully Jev-API compatible, 64k context; Cloudflare reports 209 ms (Clef) and 39 ms (Clef-flash) median latency vs Jev’s 524 ms; $0.24 per 1M input tokens on Workers AI (see Cloudflare releases Clef and Clef-flash, open-weight multimodal decision models that are Jev-API compatible)
- Perplexity follower (Oct 1, 2026): open pplx-decider-v1-27b behind a Decisions API at $0.04 per 1M input tokens; 85.71% vs Jev 84.51% on Perplexity’s 11-benchmark panel (see Perplexity launches a Decisions API and open-sources pplx-decider-v1-27b, a decision model it says edges Jev (85.71% vs 84.51%))
- Open follower (HF, created Sept 18, 2026): Convai Innovations’ Laya, an Apache-2.0 non-autoregressive ‘System 1’ decision model (421M-parameter ModernBERT-large English checkpoint; 322M mmBERT-base multilingual one) trained with ‘RLCD’ (RL with strictly proper scoring rules as rewards). Model card claims 0.766 vs Jev’s 0.727 on its typed-decisions benchmark and 7.8x lower latency (32.8 ms on a T4); ~5k Hugging Face likes by Oct 2 (self-reported numbers, not independently checked)
- Small open follower (HF, Sept 23, 2026): Supersonic Labs’ Julia 1, a 144.3M-parameter Apache-2.0 decision model fine-tuned from mmBERT-small; on its own tests 73.15% typed-decision accuracy vs 72.70% for a Jev reference (but 64% vs 87% on a Banking77 pilot); trending on Hugging Face in early Oct (huggingface.co/SupersonicLabs/Julia-1)
- Distilled open copies (HF, Sept 23–Oct 5, 2026): AutoTrust’s JEV-9B, JEV-27B (Qwen3.8-27B adapter, described as a ‘student of TypeSafe Jev 1.13’ trained largely on Jev output distributions), JEV-27B-VL (adds image input, ‘System 1’ API at POST /v1/decide; Oct 3 robot-arm and computer-use demos) and GEV-26B-Decide (Gemma 4 26B-A4B base). All Apache-2.0. AutoTrust’s own six-benchmark run puts JEV-27B 0.22 points above hosted Jev 1.13. By Oct 7, JEV-27B-VL had ~1.5M downloads and GEV-26B-Decide ~855k, among the most-trending models on Hugging Face (self-reported benchmarks; download counts may include automated pulls)
- Jev 1.14 (TypeSafe blog, Oct 14, 2026): jev-1.14.0 released under jev-preview (jev-latest to follow a week later), with small broad gains, mostly multilingual; opt-in experimental ‘danger zone’ flags for image input, a new single bounding-box
boxtype and longer context than production allows (no limit stated). jev-1.13.0 gets 60 days’ notice before deprecation. No pricing or benchmark numbers in the post - AutoTrust JEV-35B (Hugging Face, Oct 11, 2026): Apache-2.0 decision model fine-tuned from Qwen3.5-35B-A3B (~3B active), up to 256 options per pass; self-scored 59.73 on the public Decision Index 0.3 suite vs 53.64 for JEV-27B-VL, median latency 204 ms on one B200; NVFP4 version fits 32–48 GB GPUs. ~28k downloads and 771 likes by Oct 15, among Hugging Face’s top trending models (self-reported scores)
- OpenAI follower (Oct 6, 2026): OpenAI’s Decisions API entered public beta with gpt-6-luna at $0.10 per 1M input tokens (see OpenAI opens the Decisions API to all developers in public beta)
- Microsoft follower (Oct 9, 2026): Microsoft-Decision-1, a post-trained Qwen3.5-9B, at the same $0.042 per 1M input tokens with free output, in Microsoft Foundry (see Microsoft-Decision-1 decision model)
- Category spreads (TechCrunch, Oct 6): moderation startup Musubi released PolicyLM-1.7B, an open-weights decision model that applies plain-English content policies to messages in under 50 ms without retraining when policies change; TechCrunch lists Jev, OpenAI’s and Amazon’s decision models as the wave it joins. Musubi’s Filip Jankovic: ‘Product teams just want a better understanding of what’s happening on their platform’
- TechCrunch (Oct 16, 2026), on Vercel’s AI Gateway data: within two weeks 35% of paid teams on the gateway used Jev, nearly double the reach of the second model, Sonnet 5.5, the fastest two-week adoption Vercel has seen; Vercel itself replaced GPT-6 Luna with Jev for classifying harmful commands. TechCrunch notes the gateway undercounts labs with their own APIs
What happened
Jev is aimed at the many small judgments inside software (classifying, ranking, routing, reranking search results) where developers currently call a chat model and parse its text. It returns calibrated numbers in a fixed schema instead. Developers adopted it quickly: within days there were integrations, playful hacks (a 2048 player, a left-pad) and an open-source imitation built on Qwen.
Why it matters
It is a new product shape for language models, a cheap, fast “function call” for judgments, and it was widely discussed as a complement to frontier agents in the same week as Claude Opus 5.5 and GPT-6 Sol.
Sources
21 sources from 13 sites. Numbers match the chips in the text.
21 sources: 11 primary, 7 press, 3 reactions
Primary
- TypeSafe blog: Jev 1.14 (Oct 14, 2026)typesafe.ai, official
- Hugging Face: autotrust/JEV-35B model cardhuggingface.co, code
- Strands Agents blog: Introducing Strands Decider (Oct 1; Hacker News front page Oct 7, ~281 pts)strandsagents.com, official
- TypeSafe AI: Introducing System One Models & Jevtypesafe.ai, official
- Vercel: TypeSafe AI’s Jev now available on AI Gatewayvercel.com, docs
- Diogo Almeida on X: launching Jevx.com, official
- Vercel on X: Jev adopted faster than any model in AI Gateway historyx.com, official
- Hugging Face: StrandsAgents/strands-decider-2B-hobson-v19huggingface.co, code
- Cloudflare blog: Clef decision modelsblog.cloudflare.com, official
- Hugging Face: convaiinnovations/laya (Laya decision model)huggingface.co, code
- Hugging Face: autotrust/JEV-27B-VL (open distilled Jev student with vision)huggingface.co, code
Press
- TechCrunch: Jev is Vercel’s fastest-adopted AI model ever (Oct 16)techcrunch.com, press
- TechCrunch: How AI decision models could change content moderation (Oct 6)techcrunch.com, press
- WSJ: Startup TypeSafe AI’s Jev model sparks copycats, talk of LLM alternativeswsj.com, press
- AI Weekly: TypeSafe says Jev reaches 25% of Fortune 500, trillion tokens dailyaiweekly.co, press
- Forbes: Jev cuts AI decision costs 100x and Vercel, Cloudflare rushed to add itforbes.com, press
- VentureBeat: Amazon unveils Strands Decider 2B, a free open-source Jev competitorventurebeat.com, press
- The New Stack: AWS launches a local answer to TypeSafe’s Jev decision modelthenewstack.io, press
Reactions
- Nathan Flurry on X: hype-free explanation of Jev (670K views)x.com, discussion
- Jarrod Watts on X: trading bot built with Jev (1.28M views)x.com, discussion
- Simon Willison: Jev introduces a new shape of LLM - System One, aka Decision Modelssimonwillison.net, discussion
Changes
- TechCrunch on Jev’s adoption on Vercel’s AI Gateway
- Added Jev 1.14 (preview, experimental image, bounding-box and long-context flags) and AutoTrust’s JEV-35B open student
- Added Microsoft-Decision-1 as a follower
- Added the $870M a16z round
- Linked the Strands Decider blog post (HN front page Oct 7, ~281 pts)
- Added TechCrunch on Musubi’s PolicyLM-1.7B as part of the decision-model wave
Show all 16 changes
- Added AutoTrust’s distilled open Jev students (JEV-27B/-VL, GEV-26B-Decide; trending on Hugging Face) and OpenAI’s Decisions API beta
- Added Supersonic Labs Julia 1 (144M open decision model) as a further follower
- Added Perplexity pplx-decider as a further follower
- Added the WSJ interview (25% of Fortune 500, 1T tokens/day; company figures)
- Added Convai Innovations’ open Laya decision model
- Added Cloudflare Clef/Clef-flash as an open-weight, Jev-compatible follower
- Added AWS Strands Decider 2B as an open-source follower
- Linked the highest-reach community Jev posts
- Added launch-post reach, Vercel adoption data and open follow-ons; 16 related X posts archived
- Filed