Jev: The Decision Model That Went Viral

Episode 1 · Jared Palmer Open-Sources Kev, a Family of Small Judge Models Rebuilt on Qwen3.5 (2026-09-21, 5 posts)

Developer Jared Palmer open-sourced Kev, a family of small "Jev-style" judge models, and on 09-22 completed a rebuild on Qwen3.5, releasing 0.8B, 4B and 9B checkpoints under Apache-2.0 with training code and a frozen evaluation suite. The project reached Hacker News and offers a new lightweight, locally runnable option for judge models.

Confirmed

• The latest Kev is built on Qwen3.5 in 0.8B, 4B and 9B sizes, Apache-2.0 licensed, with training code, a frozen eval suite, and a fine-tuning script that runs on Modal.
• The 9B model fits in bf16 on a 32GB Mac.
• Earlier evolution: Kev-0.5B was based on Qwen2.5-0.5B with a TypeSafe-compatible API, trainable and runnable on a MacBook Pro; it was then upgraded to 0.6B/4B/8B on Qwen3 using LoRA plus a small pointer head, trainable on a single H100 in 40 minutes.
• Kev is positioned as a self-trainable, self-deployable small judge model for decision tasks.

Why it matters

• Large judge models typically require heavy compute; Kev shows small models can handle judging tasks fully locally, lowering the barrier to entry.
• Rapid iteration within a day (Qwen2.5 → Qwen3 → Qwen3.5) with training code open-sourced makes reproduction and secondary development easy for the community.

• Kev open-source decision models scale to 0.6B/4B/8B, trainable in 40 min on one H100 — TheMoonMidas · 2026-09-21
• Kev: tiny Jev-like family of decision models built on Qwen3.5 hits Hacker News — petrusenko_max · 2026-09-21
• Kev-0.5B: Tiny Open-Source Jev-Like Decision Model Built on Qwen2.5-0.5B — multiply_matrix · 2026-09-21
• Kev refactored onto Qwen3.5: open-source decision models now at 0.8B, 4B and 9B — alexcovo_eth · 2026-09-22
• Kev: open-source 0.8B/4B/9B judge models on Qwen3.5, the 9B fits a 32GB Mac — khiladi1729 · 2026-09-22

Episode 2 · Classification Model Jev Goes Viral for Fast, Cheap AI Judgments (2026-09-24, 2 posts)

TypeSafe's classification model Jev, released on September 15, performs fast and cheap judgment tasks rather than generation. A viral demo showed it processing 724 ads from 37 brands in 40 seconds with over 8,700 classifications for just $0.09.

• Jev classifies 724 live ads from 37 brands in 40s for $0.09 in tokens — sibraan_ · 2026-09-24
• Jev, a classification-not-generation model, goes viral with 1M-view intent demos — danshipper · 2026-09-25

Episode 3 · Jev, a judgment-only model, takes off on Vercel AI Gateway (2026-09-25, 8 posts)

Stealth frontier lab TypeSafe AI launched Jev on September 15, and this week it went free on Vercel AI Gateway. Jev is a "System 1" fast structured decision model: it generates no text, instead taking shared state and typed questions and returning narrow, code-executable decisions like booleans and probabilities, under the philosophy "code owns the workflow, AI handles narrow decisions." Official figures claim up to 200x faster and 400x cheaper than LLMs on equivalent classification tasks; within 24 hours nearly 13% of paid teams adopted it, and about 140,000 developers joined the beta in 36 hours.

Confirmed

• Released by TypeSafe AI, Jev is free on Vercel AI Gateway with 32K context and promotional pricing through September 25, 2026 (@JohnPhamous)
• Jev does classification, scoring and routing only — e.g. judging whether an email is urgent or whether text sounds AI-written (@rseroter, @adnanhashmi, @every)
• Official claims: up to 200x faster and 400x cheaper than LLMs; pricing is $0.042 per million input tokens with free output (@multiplymatrix relaying)
• In a small independent test, a general-purpose model spent 910 output tokens per yes/no answer (@rseroter)
• Input is 238x cheaper with free output; in a 10K-token state × 1000 decisions scenario, Astra costs $100 versus $0.42 for Jev, with architecture rather than unit price being the real saver (@multiplymatrix)
• Per @创业邦: 13% of paid teams adopted it within 24 hours, over twice the adoption rate of comparable new models; 140,000 developers joined the beta in 36 hours
• Per @every, Jev takes text or structured data and outputs a probability suited to high-frequency calls; developer @hammermt built Jevgram on it for per-question text scoring and live coding-agent correction
• Delip Rao notes Jev offers calibrated confidence most LLMs lack, making it a natural fit for cascaded decision architectures and "fast/slow" dual-system designs, e.g. replacing Flash as the fast system
• Every argues the rise of judgment models signals classification tasks driving "intent-driven" software (@danshipper relaying)

Why it matters

• Jev represents a new model type: stripping LLMs out of decision points, letting code own workflows while AI handles narrow decisions, potentially cutting architectural costs of large-scale decision-heavy applications
• Rapid adoption plus real integrations like Jevgram suggest genuine production demand rather than hype
• Calibrated confidence plus ultra-low latency provides a building block for cascaded/dual-system architectures, possibly spawning new AI application paradigms

• TypeSafe's Jev: A Model That Can't Write but Decides, Claiming 200x Speed and 400x Cost Savings — rseroter · 2026-09-25
• TypeSafe AI's Jev evaluation model goes free on Vercel AI Gateway — JohnPhamous · 2026-09-25
• Jev is 238x cheaper than Astra on input — architecture, not price, saves the money — multiply_matrix · 2026-09-25
• TypeSafe's decision model Jev claims 200x speed, 400x cost savings on agent routing — adnan_hashmi · 2026-09-26
• Jev's calibrated confidence enables fast-slow cascade judging, 1-2 orders cheaper — deliprao · 2026-09-26
• No Chatter, Just Judgment: Jev Model Wins 13% of Paid Teams in 24 Hours — 创业邦 · 2026-09-26
• Jev: fast cheap probability models let coding agents self-edit while writing, not after — every · 2026-09-26
• TypeSafe's Jev classification model goes viral, ushering in intent-based software — danshipper · 2026-09-27

Episode 4 · TypeSafe AI Launches Decision Model Jev, Wins LangChain Endorsement (2026-09-26, 5 posts)

TypeSafe AI launched a new model, Jev, last week and quickly won high-profile backing from LangChain: founder Harrison Chase publicly recommended the Jev + LangGraph pairing, and the team published a production-focused tutorial.

Confirmed

• TypeSafe AI released Jev, a decision-oriented model: instead of generating text, it returns typed decisions that code can execute directly — what the company calls an "AI-powered decision model."
• LangChain founder Harrison Chase (09-26) publicly endorsed combining Jev with LangGraph. He argues agents should be modeled as "complex systems with AI embedded," that LangGraph is the best way to build such systems, and that Jev fits this approach well.
• LangChain team member huntlovell amplified TypeSafe AI's post and released the tutorial "Building Prod with Jev and LangGraph" on building production-grade AI systems with the two.
• Josh Rosen wrote a piece demonstrating how to use Jev.

Why it matters

Jev represents an output paradigm distinct from mainstream LLMs: skipping the natural-language middle layer and directly producing typed, executable decisions. The rationale is that most LLM calls inside agents are effectively decisions — which tool to call, which model to use, whether an action is safe, whether a task needs escalation — and specializing for this removes unnecessary text generation. Combined with an orchestration framework like LangGraph, it could simplify the model-output-to-code-execution pipeline in production systems. The LangChain founder's personal endorsement plus an official tutorial signals this pairing is worth watching.

• LangChain's founder: Jev + LangGraph is the combo for modeling agents as complex systems — hwchase17 · 2026-09-26
• LangGraph maps well onto Jev, TypeSafe's new decision-making AI model — hwchase17 · 2026-09-26
• TypeSafe launches Jev, a decision-making model that outputs executable decisions instead of text — hwchase17 · 2026-09-26
• LangChain on Building Prod Systems with Jev and LangGraph — LangChain · 2026-09-26
• Jev, a decision-making model that skips text generation, integrates with LangGraph — airesearch12 · 2026-09-26