AutoTrust releases open-weights JEV-35B and NVFP4

AutoTrust’s open-weights JEV-35B scores about 59.7 on Decision Index 0.3; its 25.6 GB NVFP4 build scores 59.18.

AutoTrust AI published autotrust/JEV-35B, an open-weights System 1 decision model based on Qwen3.5-35B-A3B with 35 billion parameters and about 3 billion active per token, and also released the quantized autotrust/JEV-35B-NVFP4 build. On public Decision Index 0.3, sources disagree slightly on the full-precision score, citing 59.73 or 59.75, compared with 53.64 for JEV-27B-VL, while both agree the NVFP4 score is 59.18. The base model completed all 140,178 scoreable requests with zero errors; an approximate vision rebuild scored 71.94, and the NVFP4 vision score is reported 0.85 points lower, with 96.6% of text answers unchanged versus bf16. Median latency was 204 ms on one NVIDIA B200, and the quantized model is described as having similar single-request latency on one B200. NVFP4 quantizes 10,240 routed expert MLPs while leaving attention, routers, the vision tower, and the decision head in bf16, reducing the download from 71.9 GB to 25.6 GB and vLLM weight memory from 66.5 GiB to 23.3 GiB, versus about 23 GiB in the earlier report. AutoTrust says the model is not affiliated with TypeSafe AI’s hosted TypeSafe Jev 1.13.

• On 2026-10-11, AutoTrust published autotrust/JEV-35B (Hugging Face trending #28) and autotrust/JEV-35B-NVFP4 (trending #21).
• JEV-35B is an open-weights System 1 mixture-of-experts model on Qwen3.5-35B-A3B: 35 billion parameters, about 3 billion active per token, and up to 256 options.
• Decision Index 0.3 is reported as 59.73 in one source and 59.75 for bf16 in the other, versus 53.64 for JEV-27B-VL; NVFP4 scores 59.18.
• The base model completed all 140,178 scoreable requests with zero errors; cited gains include PhishNChips +42.2, HoVer +28.4, and Habermas Machine +28.1.
• An approximate vision rebuild scored 71.94; the NVFP4 vision score is 0.85 points lower, and 96.6% of text answers match bf16.
• NVFP4 quantizes 10,240 routed expert MLPs and keeps attention, routers, the vision tower, and the decision head in bf16, cutting download size from 71.9 GB to 25.6 GB and vLLM weight memory from 66.5 GiB to 23.3 GiB (earlier report: about…
• Median latency was 204 ms on one NVIDIA B200; NVFP4 single-request latency is described as similar. AutoTrust says it is not affiliated with TypeSafe AI’s hosted TypeSafe Jev 1.13.