AutoTrust AI releases JEV-27B, an open decision model mirroring closed rival Jev

Overview

Updated Sep 29

A Singapore lab says it reproduced the decision behavior of a $10-billion-valued rival with about nine hours of training on a single GPU. AutoTrust AI released JEV-27B on September 29, an open-weights model that answers yes/no, multiple-choice, and 0-to-5 rating questions in one forward pass and returns a calibrated probability for each option.

The target is TypeSafe AI's closed Jev 1.13, the System 1 decision model behind TypeSafe's reported $1-billion raise at a valuation above $10 billion. AutoTrust says JEV-27B's probability distributions sit within a mean divergence of about 0.017 of Jev's, where 0 means identical, and that on six public decision benchmarks JEV-27B averages 84.07 against Jev's 83.85. The decision block is small enough that it leaves the frozen Qwen backbone's reasoning untouched.

Why it matters

If open models match closed decision engines, enterprises can self-host agent decisions — undercutting a $10-billion-valued closed rival.

Questions about this story

Free account needed to ask — your question is kept and asked for you right after sign-up. Answers are public.

No questions yet — be the first to ask.

Key Indicators

108.9M

Trained decision-block parameters

Roughly 0.4% of the frozen 27-billion-parameter Qwen backbone.

9.2 hours

Training time on one NVIDIA B200

Reported cost to train the decision block.

84.07

Average score on six decision benchmarks

AutoTrust's runs put JEV-27B at 84.07 vs 83.85 for TypeSafe Jev 1.13.

78.0%

HumanEval score, unchanged from base model

All 164 completions byte-identical with the decision block off.

Voices

Curated perspectives — historical figures and your fellow readers.

Ever wondered what historical figures would say about today's headlines?

Organizations Involved

AutoTrust AI

AI research lab

Released JEV-27B under Apache-2.0

Singapore lab focused on open-weights decision models for AI agents.

TypeSafe AI

AI company

Maker of closed Jev 1.13; reportedly raising $1B+ at $10B+ valuation

Company behind Jev, the closed System 1 decision model for agent workflows.

Alibaba (Qwen / Tongyi Lab)

Technology company

Provided the frozen Qwen3.8-27B backbone

Alibaba's open Qwen models form the System 2 backbone of JEV-27B.

Timeline

1 event Latest: September 29th, 2026 · 1 week ago

September 2026
AutoTrust AI releases JEV-27B
Latest Product Launch

Open-weights model adds calibrated System 1 decisions to a frozen Qwen3.8-27B backbone; 108.9M-parameter block trained in 9.2 hours on one B200.
September 29th, 2026

Scenarios

JEV-27B lands named production deployments

Likely Resolves by Q2 2027

Discussed by: AutoTrust AI's demo reel and the Hugging Face release coverage

AutoTrust's launch reel shows JEV-27B piloting agent workflows: playing Doom, steering a simulated drone in MuJoCo, verifying Google Flights results, navigating Wikipedia, and routing support tickets. A named enterprise adopting the model in production would confirm the open path. The main obstacle is AutoTrust's own warning about blind spots in multi-hop reasoning, arithmetic, and dates, plus English-centric training data.

TypeSafe closes $1B+ round at a $10B+ valuation

Likely Resolves by Q1 2027

Discussed by: The reported talks covered in the Medium analysis of Jev's valuation

TypeSafe AI wraps its reported round at a valuation above $10 billion, about 50 times its seed valuation from days earlier. That would signal investors still place the premium on a closed, supported model with a public leaderboard lead, even with an open copy that reproduces its decisions.

Independent benchmarks confirm JEV-27B's parity with Jev

Uncertain Resolves by Q1 2027

Discussed by: The KL divergence measurements AutoTrust reported

A third party reproduces AutoTrust's claim: JEV-27B's probability distributions sit within a mean KL of about 0.017 of Jev 1.13's, and JEV-27B scores higher on four of six public decision benchmarks. Independent confirmation would validate the distillation approach. A larger measured gap would dent the parity claim.

Historical Context

3 moments from history that rhyme with this story — and how they unfolded.

March 2023

Alpaca's cheap distillation of ChatGPT (2023)

Stanford researchers fine-tuned Meta's LLaMA on 52,000 self-instruct examples for under $600, reproducing much of ChatGPT's behavior in an open-weights model.

Then

Sparked a wave of open instruction-tuning; the Alpaca approach was copied across the community within weeks.

Now

Established that closed-model behavior could be captured cheaply and openly, accelerating the open-weights ecosystem.

Why this matters now

JEV-27B repeats the pattern one step later: reproducing a closed decision model's outputs with a small, cheaply trained block instead of a full chat model.

February 2023

Meta releases Llama (2023)

Meta released Llama, its 7B-to-65B parameter language models, openly to researchers. Within days they leaked publicly, and the open-weights era began.

Then

Researchers and startups built fine-tunes and derivatives at a pace closed labs couldn't match.

Now

Open-weights models became the default for self-hosted deployment, pushing closed labs to compete on safety and support.

Why this matters now

JEV-27B continues that trajectory, now targeting the decision layer of agent workflows rather than chat.

August 2022

Stable Diffusion (2022)

Stability AI released Stable Diffusion, an open-weights image model that rivaled closed systems like DALL-E and Midjourney in quality while running on consumer GPUs.

Then

Open image generation spread rapidly; closed labs were forced to justify their prices.

Now

Self-hostable open models became a durable option for image generation, especially where data privacy mattered.

Why this matters now

Like Stable Diffusion, JEV-27B offers a self-hosted open alternative to a closed, high-value model, targeting enterprises that want to keep agent decisions in-house.