AutoTrust releases GEV-26B-Decide and NVFP4 variant

AutoTrust released open GEV-26B-Decide scoring 62.48 on Decision Index 0.2.1, then an NVFP4 build cutting vLLM memory to about 17 GiB.

AutoTrust released autotrust/GEV-26B-Decide, open weights previously named JEV-Gemma4-26B-A4B and built on Google's Gemma-4-26B-A4B-it (26 billion parameters, about 4 billion active). It is not a TypeSafe product. On Decision Index 0.2.1 it scores 62.48 balanced skill, above TypeSafe Jev 1.13 at 57.91, with Tools and Automation at 0.697 and Arts and Taste at 0.415. System 1 returns calibrated option probabilities in one pass (about 45 ms on a B200) and adaptive thinking invokes Gemma-4 reasoning below 0.8 confidence. In demos it finished 95% of 60 browser tasks at about 85 ms per click, but only 40% of 20 robot-arm pick-and-place scenes. On 2026-10-05 AutoTrust also published GEV-26B-Decide-NVFP4, quantizing the 3,840 routed expert MLPs to NVFP4 while leaving attention, routers, the vision tower, and decision heads in bf16: download size drops from 49.5 GB to 18 GB and vLLM weight memory from 51.1 GiB to 17.1 GiB, with System 1 GPQA Diamond moving from 43.9% to 44.9% and HLE text multiple choice from 8.4% to 9.7%, within run-to-run noise. Sources agree on the bf16 Decision Index score of 62.48; the later report does not restate the robot-arm timing given as 61 ms per step earlier.

• AutoTrust published open weights autotrust/GEV-26B-Decide (earlier name JEV-Gemma4-26B-A4B) on Google Gemma-4-26B-A4B-it: 26B parameters, about 4B active; not a TypeSafe product.
• Decision Index 0.2.1 balanced skill is 62.48, above TypeSafe Jev 1.13 at 57.91; area scores include Tools and Automation 0.697 and Arts and Taste 0.415.
• System 1 returns calibrated option probabilities in one pass (about 45 ms on a B200); adaptive thinking uses Gemma-4 reasoning below 0.8 confidence.
• Demos: 95% of 60 browser tasks at about 85 ms per click, and 40% of 20 robot-arm pick-and-place scenes (61 ms per step in the first report).
• On 2026-10-05, autotrust/GEV-26B-Decide-NVFP4 quantizes 3,840 routed expert MLPs to NVIDIA NVFP4; attention, routers, vision tower, and decision heads stay bf16.
• Download size falls from 49.5 GB to 18 GB; vLLM weight memory from 51.1 GiB to 17.1 GiB. Native W4A4 on Blackwell; Marlin W4A16 on Hopper and Ampere.
• System 1 on the NVFP4 build: GPQA Diamond 43.9% to 44.9% and HLE text multiple choice 8.4% to 9.7%, described as within run-to-run noise. The bf16 model still scores 62.48 on Decision Index 0.2.1 with adaptive thinking.