TypeSafe
TypeSafe sells Jev, a "System One" model that makes typed decisions instead of writing text. It is fast and cheap enough to call inside ordinary code paths, and it cannot generate prose.
What Jev is
Diogo Almeida, a ChatGPT co-inventor, launched TypeSafe and Jev on 2026-09-15 after two years building a new training method called RLCD. The launch claims Jev is 20 to 200 times faster and 40 to 400 times cheaper than an LLM for decisions. Input costs $0.042 per million tokens and output tokens are free. The trade is stated plainly: Jev cannot generate text. It answers typed questions (pick one of N, yes or no) with a probability for each answer. The launch demos race between Wikipedia pages by choosing among hundreds of links at about ten calls a second, roughly $7 an hour. Source: X/@CompleteSkeptic thread, 2026-09-15; Source: TypeSafe blog
Where it fits
Use Jev where code needs a judgment and an LLM would be the expensive part: classify, route, rank, choose an action from a fixed menu. Code keeps the thresholds; Jev returns the probabilities. Keep LLMs for anything that must write.
Kevin's projects use it this way:
- private page sends email metadata to Jev's System One endpoint, asks five typed questions per message, and applies its own thresholds, for about five cents per thousand emails.
- private page lets Jev pick one action from a menu of visible controls, failing closed above 255 candidates.
The typesafe:typesafe-ai skill holds the API patterns.
Harnesses built on Jev
A Minecraft speedrun shows the split between a fast decider and a slow planner. Jev controls the player (WASD, jump, click, mouse) and makes each move; GPT-6 Astra plans and sends instructions to Jev asynchronously. When a run failed, Astra wrote a new skill as an .mjs file for that situation, until the agent beat the Ender Dragon in 8 minutes 43 seconds for under a dollar ($0.01 of Jev, $0.96 of Astra). The author notes a preselected seed and that it was not good enough for real-time combat on Normal. Source: X/@rronak_ thread, 2026-09-20; Source: rmalde/minecraft-agent
The keel builder's guide puts Jev at the decision layer of a coding workspace: a selector (local Laya, Jev, or a normal LLM) chooses, and the host checks the choice before the coding agent runs the task. Source: X/@Av1dlive, 2026-09-23; Source: codejunkie99/keel
Local typed decisions (Laya)
Laya is an open-weight typed-decision model in the same family as Jev: it scores answers instead of writing text. laya-mlx ports it to Apple Silicon at about 13 ms per short decision on an M3 Max with under 1 GB of memory, fast enough to play Snake at 60 decisions a second; laya-coreml runs it on the Neural Engine at about 5 ms with up to 64% less power. Use Laya when a decision must stay on the device or run in a tight loop; use Jev when hosted quality matters more. Version checked: laya-mlx v0.6.0 (2026-10-29), Apache-2.0. Source: X/@mizorewww thread, 2026-09-20; Source: mizorewww/laya-mlx; Source: gh release list, 2026-11-03
Generative UI at decision speed
Chris Tate paired json-render with Jev so generated UI renders in milliseconds: Jev chooses from your own components and actions instead of an LLM writing them. The experiment exposed experimental_composeSpec and experimental_createEvaluator in json-render. The same shape as private page: a fixed menu, a typed choice, code renders the result. Version checked: json-render v0.21.4 (2026-11-01). Source: X/@ctatedev thread, 2026-09-18; Source: vercel-labs/json-render
Other typed-decision models
Liquid AI's LFM3 family includes two decision models, d1-2B and d1-Omni-2B, that return typed outputs with calibrated probabilities in one pass; the Omni version also takes audio. The rest of the family is small on-device models: LFM3-1.7B for local text agents with 256K context, LFM3-VL-2B for video and spatial understanding, LFM3-Audio-2B for voice and tool calls in 18 languages, and an omni model. A tri-hybrid architecture (gated short convolutions, sliding-window and global attention) was searched for 14 device and backend targets, and each model can be fine-tuned on one GPU. Source: X/@liquidai thread, 2026-10-27; Source: Hugging Face LiquidAI/LFM3-1.7B
Typed decisions now come in three forms: hosted (Jev), open-weight on Apple Silicon (Laya) and open-weight on any device (LFM3 d1). Pick by where the decision must run.
Open models behind the same API
SGLang added /v1/decisions, which turns any chat model or VLM into a decision model: you send an input and typed questions, and each answer's probabilities are read from the model's next-token scores at the answer position. Nothing is generated or parsed, and no special checkpoint is needed. SGLang also added /v1/systemone so open models can stand in for Jev behind the TypeSafe SDK. Their demo turned Qwen3.8-27B into a multimodal decision model that beat Pokémon FireRed's Elite Four and Champion with decisions under 100 ms. Shipped in v0.5.21. Source: X/@sgl_project thread, 2026-09-30; Source: SGLang docs, decision models
A language with Jev inside
Probably is a toy language that makes Jev the if: feels asks a question, match routes between descriptions, and while runs until something stops feeling true. Jev decides, an LLM writes, and a short program ties them together. It is a toy, but it shows the right division of labor. Source: X/@southpolesteve, 2026-09-18; Source: southpolesteve/probably
Adoption
Vercel reported Jev as the fastest-adopted model in AI Gateway history: about 13% of teams on day one, twice the GPT-5.6 family and six times Fable 5.1. Source: X/@vercel, 2026-09-18; Source: Vercel blog
Distribution
TypeSafe came out of stealth on 2026-09-15, and Jev reached Vercel's AI Gateway two days later as typesafe-ai/jev, called through evaluate({ model, state, questions }) with typed Choice, Score and Boolean answers. Source: X/@typesafeai, 2026-09-15; Source: Vercel changelog
Timeline
- 2026-11-03 | Page created from the Jev launch bookmark and the two projects that use it. Source: X/@CompleteSkeptic, 2026-09-15
- 2026-09-15 | TypeSafe launched Jev publicly. Source: X/@CompleteSkeptic, 2026-09-15