Jev directory
CommunityJev Models & Reimplementations
AutoTrust JEV-35B GGUF
prithivMLmods' GGUF builds of autotrust/JEV-35B, the Qwen3.5-35B-A3B MoE decision model that returns calibrated choice, noul and score answers over text and images; files run from BF16 to Q3_K_L for llama.cpp.
- Reading time
- 2 min
- Published by
- Community
- Author
- prithivMLmods
- Last verified
- Oct 2026
- Capabilities
- Open sourceSelf-hostedRuns locally
Highlights
- GGUF builds of autotrust/JEV-35B from BF16 at 71.1 GB down to Q3_K_L at 18.6 GB, with Q4_K_M, Q5_K_M and Q6_K marked recommended, plus a 903 MB mmproj.
- The card reports Decision Index 0.3 of 59.73 in AutoTrust's own scoring against 53.64 for JEV-27B-VL, leading all five areas and 24 of 37 benchmarks.
- Median latency 204 ms on one B200 versus 271 ms for JEV-27B-VL; the decision head sits in the lm_head rows, so vLLM loads one checkpoint with no LoRA.
- ERQA 51.3% in one pass with no robot data; 95% computer use with element text, 38% with numbered boxes; the card recommends JEV-27B-VL for game play.
- MTP speculative decoding is enabled, and an NVFP4 build at 25.6 GB scores 59.18 and fits on one 32-48 GB GPU; Apache-2.0.
Watch out
Card reviewed 2026-10-17; the 59.73 Decision Index and 71.94 vision figures are AutoTrust's own scoring with its kit, not board entries, the vision comparison is a private rebuild of unpublished benchmarks, Knowledge and Reasoning stay weak (GPQA 0.360, HLE at chance), and ERQA carries roughly plus or minus 5 points at 400 questions.
Questions about this entry
Short answers drawn from the linked project or page.
prithivMLmods' llama.cpp GGUF conversion of autotrust/JEV-35B, AutoTrust's System 1 decision model. It answers typed noul, choice and score questions over text and images and returns a calibrated probability for every option in one forward pass.
The card reports Decision Index 0.3 of 59.73 versus 53.64 for JEV-27B-VL, a 71.94 versus 71.67 vision tie, and a 204 ms versus 271 ms median latency on one B200. All of these are AutoTrust's own numbers.
The card marks Q4_K_M (21.7 GB), Q5_K_M (25.3 GB) and Q6_K (29.2 GB) as recommended, with BF16 at 71.1 GB for highest quality and Q3_K_L at 18.6 GB for low RAM. Pair any file with the 903 MB mmproj-bf16 for vision.
No. JEV-35B is AutoTrust's open-weights decision model, a fine-tune of Qwen3.5-35B-A3B, distinct from TypeSafe's hosted Jev 1.13. This repository only repackages it as GGUF files for llama.cpp runtimes.
The decision head is merged into the lm_head rows of the slot tokens, so vLLM loads one plain checkpoint with no LoRA adapter. It is a decision-only model: no chat, no text generation and no thinking mode.
The card concedes it trails JEV-27B-VL on GSM8K, ANLI, BBH and POP909-CL, that Knowledge and Reasoning stay weak with GPQA at 0.360 and HLE at chance, and that for game play it recommends JEV-27B-VL instead.
Six files: BF16, Q3_K_L, Q4_K_M, Q5_K_M, Q6_K and the mmproj-bf16 vision projector. MTP speculative decoding is enabled in the weights. The card links llama.cpp as the runtime but does not include an example command.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
LinkedIn: What multimodal decision models actually change, and what they do not.
LinkedInJevDirectory.org
X: Standard One SH (StandardThinking)
Hacker News: Decision Index 0.2.1 scores open reproductions of Jev on a frozen suite.
jackbravo4 pointsDecision Index 0.2.1 – 70 open reproductions of TypeSafe Jev's Decision ModelRead the thread on Hacker NewsHacker News: Pointer to a DiffusionGemma plus vLLM tutorial for a Jev-shaped classifier.
1 pointHow to build a Jev-style classifier with DiffusionGemma and vLLMRead the thread on Hacker NewsMore Jev resources like this
A 4-bit W4A8 AWQ plus GPTQ build of AutoTrust's JEV-27B-VL from AI Outfitters that keeps the rank-32 LoRA decision head and a bf16 vision tower, served by vLLM on a single 24 GB GPU.
Jev Models & Reimplementations#community#open-models#local
Community
A community Q8_0 GGUF of Mapika/decider-0.8b, the smallest System One typed-decision model: one forward pass returns calibrated Choice, Score and noul answers, reporting 0.405 on JevBench hard-public at 920 MB resident.
Jev Models & Reimplementations#community#open-models#local
Community
A 4-bit AWQ plus GPTQ build of AutoTrust's JEV-35B from AI Outfitters: INT4 weights with FP8 and INT8 activations, calibrated on all 256 experts, and served by vLLM on a single 24 GB GPU.
Jev Models & Reimplementations#community#open-models#local
Community
Mapika's NVFP4 fine-tune of Gemma-4-31B-it answers typed Choice, Noul and Score questions with calibrated probabilities and no generated text, self-reporting a 64.27 chance-corrected public Decision Index and 39 ms median latency on a B300.
Jev Models & Reimplementations#community#open-models#local
Community
AutoTrust's JEV-9B on ModelScope: a Blocks of Experts System 1 + System 2 model fine-tuned from Qwen3.5-9B, Apache-2.0, about 8.95 GB of BF16 safetensors in five shards, taking images since 3 October 2026, with a GuruSearch live demo.
Jev Models & Reimplementations#community#open-models#local
Community
akhilaaa3's merged multimodal decision classifier on Gemma 4 12B IT: probabilities for text, image, audio or video options, 87.57% on DecisionBench Medium and MMAU 63.10%, Apache-2.0.
Jev Models & Reimplementations#community#open-models#local
Community
From the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.