CommunityJev Models & Reimplementations
AutoTrust JEV-35B NVFP4
The NVFP4 build of AutoTrust's JEV-35B: routed experts quantized to 4-bit cut the download to 25.6 GB and GPU weights to 23.3 GiB, keeping 96.6% of bf16's Decision Index answers.
- Published by
- Community
- Author
- AutoTrust AI Lab
- Last verified
- Oct 2026
- Capabilities
- Open sourceSelf-hostedRuns locally
Highlights
- Only the 10,240 routed expert MLPs (40 layers x 256 experts) are NVFP4; attention, routers, vision tower and the decision head stay bf16.
- Download 25.6 GB and 23.3 GiB of GPU weights, a 65% cut from bf16's 66.5 GiB, so one 32-48 GB GPU is enough.
- Decision Index 0.3 public 59.18 against bf16's 59.73, with 96.6% identical top-1 answers across 308,559 answers.
- Median single-request latency 227 ms text and 371 ms image on one B200, essentially bf16's 215/378 ms.
- Uses the ModelOpt NVFP4 layout, so vLLM loads it without extra flags; native W4A4 was tested on B200 only.
Watch out
Quantization costs are small but real: -0.57 on the Decision Index, -0.85 on the vision rebuild (paired bootstrap), and the robot-arm demo drops from 15 to 13 of 20 scenes, within noise by sign test (p = 0.63). Native W4A4 was tested only on B200; the SM12x and Hopper rows are untested, and the card marks the games results noisy.
Questions about this entry
Short answers drawn from the linked project or page.
Only the 10,240 routed expert MLPs across 40 layers and 256 experts each. Attention, linear attention, the shared expert, routers, vision tower, MTP layer, lm_head, System 1 adapter, decision head and temperatures all stay unchanged.
The download is 25.6 GB against 71.9 GB in bf16, and GPU weights take 23.3 GiB instead of 66.5 GiB, a 65% cut that fits one 32-48 GB GPU. Latency is unchanged: 227 ms text and 371 ms image medians against bf16's 215/378.
The card reports -0.57 on the Decision Index public suite (59.18 versus 59.73) and -0.85 on the vision rebuild (71.23 versus 72.08), with 96.6% of Decision Index answers and 95.2% of vision answers identical to bf16.
B200, B300 and GB200 run native W4A4 kernels and were tested on B200. RTX PRO 6000, RTX 5090 and DGX Spark use CUTLASS FP4 kernels untested here; H100, H200 and A100 fall back to Marlin W4A16, also untested.
No. The checkpoint uses the ModelOpt NVFP4 layout with quant_method modelopt, so vLLM reads the quantization from config.json and hf_quant_config.json. serve.sh and the API are exactly those of JEV-35B.
The robot-arm pick-and-place demo dropped from 15 to 13 of 20 scenes, but 13 of 13 grasped cubes still reached the tray and the change is within noise (sign test p = 0.63). Computer use held at 95% with numbered boxes plus element text.
The card marks them as noisy single runs at seed 0: 2048's largest tile reached 64 against bf16's 32, Snake 5 food against 11, Quick Draw 85.9% against 86.3%, and chess mate-in-one 45% against 47%.
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
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 NewsHacker News: A one-token logprob wrapper that also classifies webcam frames.
allanrbo151 pointsA single function Jev-like wrapper for LLMs, including vision modelsRead the thread on Hacker NewsMore Jev resources like this
AutoTrust's bf16 System 1 decision model on Qwen3.5-35B-A3B with about 3B active parameters: Decision Index 0.3 public 59.73, ERQA 51.3% in one pass with zero robot training, 66.5 GiB of weights.
Jev Models & Reimplementations#community#open-models#local
Community
Vision-capable GGUF port of autotrust/JEV-27B-VL: BF16 through Q3_K_M with a separate mmproj file, running the calibrated System 1 typed-decision interface and a 27B System 2 through llama.cpp.
Jev Models & Reimplementations#community#open-models#local
Community
Reza2kn's llama.cpp Q4_K_M build of Jev-Omni: a 6.87 GiB multimodal decision backbone plus projector and FP32 head, measured at 8.98 GiB VRAM on an RTX 5080 Laptop GPU.
Jev Models & Reimplementations#community#open-models#local
Community
AutoTrust's combined System 1 and System 2 bundle on Gemma-4-26B-A4B-it: Decision Index 0.2.1 of 58.05, a 24-slot head with noul, choice and score, and the base chat model left bit-identical.
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
Ruiruiz30's Apple Silicon MLX conversion of Jev-Omni: 4-bit affine language weights on a Mac mini M4 16GB, about 7 GB peak Metal memory and roughly 963 ms text decisions, Apache-2.0.
Jev Models & Reimplementations#community#open-models#local
Community
From the Jev guides
Original write-ups on what this entry is for.
Training recipes, looped architectures, and language ports: what the open side of the typed-decision class shipped this month, and how to pick by runtime and calibration instead of the leaderboard.
Jev Models & Reimplementations#open-models#training#local
Read article
Taking a decision checkpoint out of Python: what the GGUF and browser WebGPU exports preserve, what they change, and the parity evidence the model cards publish.
Jev Tools & CLIs#open-models#local#onnx
Read article
Interface, size, license, claims — in that order. A selection guide across the system-one catalog, from a 706K form scorer to 27B Jev students.
Jev Practices & Patterns#open-models#evaluation#benchmarks
Read article
From the Jev community
Posts from builders shipping with Jev right now.