AutoTrust AI Launches 45ms GEV-26B Decision Model

AutoTrust AI Lab has introduced GEV-26B-Decide, a specialized model built upon Google's Gemma-4-26B-A4B-it base architecture. Designed as a calibrated decision API, the system maps queries directly to structured probability distributions across predefined options, eliminating the need to parse text strings like "A" or "B." The architecture features two operational modes: a rapid System 1 path that bypasses token generation to return decisions in roughly 45 ms via a single forward pass, and an adaptive System 2 reasoning mode.

When the leading choice in the fast path falls under a confidence threshold of 0.8, the system automatically triggers step-by-step thinking using a set think_budget token cap. Once deep reasoning completes, GEV-26B-Decide blends the initial and post-thinking predictions using an equal weighting formula (p = 0.5p1 + 0.5p2) to preserve calibrated confidence. This dual-path approach drives substantial benchmark gains, elevating GPQA Diamond scores from 42.9 to 78.6, BBH from 75.0 to 92.0, and MMLU-Pro from 65.0 to 84.6.

The model handles a wide variety of classification needs, supporting binary yes/no queries, 0-5 rating scales, and multiple-choice selections up to a 256-way choice. It natively processes both text and image inputs across a 256K token context window. In practical application tests, a web browser agent driven by the model completed 95% of multi-step tasks, averaging approximately 85 ms per click when running on a single B200 GPU.

For developers and system architects, GEV-26B-Decide simplifies automated routing and decision pipelines by providing reliable confidence metrics. The system serves responses through a vLLM engine equipped with a dedicated POST /v1/decide route alongside standard OpenAI-compatible endpoints. AutoTrust AI Lab has released the adapter, decision head, and associated calibration files under an open-source Apache-2.0 license, while the base Gemma model remains governed by its original terms.