Jev directory
CommunityJev Models & Reimplementations
JEV-27B-VL AWQ INT4 W4A8
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.
- Reading time
- 2 min
- Published by
- Community
- Author
- AI Outfitters
- Last verified
- Oct 2026
Used inJev for Software & DevTools
- Capabilities
- Open sourceSelf-hostedRuns locally
Highlights
- Every text Linear, including the GatedDeltaNet projections, is int4 group 128 symmetric with int8 per-token activations on Marlin W4A8 kernels.
- The rank-32 LoRA decision head stays a LoRA exactly as the base ships; the bf16 vision tower is spliced back after quantization.
- Self-reported on 500 held-out rows: KL mean 0.0147 versus 0.0143 for the deployed W4A16 build, and choice top-1 0.825 versus 0.851.
- Cold prefill of a 5,098-token decision takes 1.76 s (0.345 ms/token), called parity with the W4A16 build because the workload is bandwidth-bound.
- Image decisions verified with red circle to circle at p=0.99999; served by vLLM on one 24 GB GPU with the provider's own serve scripts.
Quickstart
bash
hf download aioutfitters/JEV-27B-VL-AWQ-INT4-W4A8 --local-dir JEV-27B-VL-AWQ-INT4-W4A8Watch out
Card reviewed 2026-10-17; all validation numbers are the publisher's own runs against the provider's 500-row held-out corpus through its /v1/decide shim, the baseline is the W4A16 build rather than the unquantized checkpoint, and GPU execution claims rest on a single RTX 4090.
Questions about this entry
Short answers drawn from the linked project or page.
A 4-bit quantization of AutoTrust's JEV-27B-VL built by AI Outfitters with llm-compressor. Every text linear is int4 with int8 activations, while the LoRA decision head and the vision tower keep their original precision.
The base checkpoint's LoRA decision head is kept as a LoRA (rank 32) exactly as it ships, and the lm_head stays bf16; the bf16 vision tower is untouched and spliced back after quantization.
On the same 500 seeded held-out rows: KL mean 0.0147 versus 0.0143, choice KL 0.0385 versus 0.0374, top-1 choice 0.825 versus 0.851 and the billing example at p=0.9963 versus 0.9976.
Yes: the card reports image decisions verified with red circle to circle at p=0.99999, so the bf16 vision tower is intact after quantization.
Cold prefill of a 5,098-token decision is 1.76 s (0.345 ms/token), which the card calls parity with the deployed W4A16 build at 0.350 ms/token because the workload is bandwidth-bound rather than GEMM-bound.
Download the repository, then start vLLM with --enable-lora --max-lora-rank 32 and the adapter as a lora module, plus --kv-cache-dtype fp8 --gpu-memory-utilization 0.95 --max-num-seqs 8 on a 24 GB card; the decision_head folder carries the bias and temperatures.
Apache-2.0, the same as autotrust/JEV-27B-VL. The quantization, calibration and validation are AI Outfitters' own work; the upstream model comes from AutoTrust and is unrelated to TypeSafe's hosted Jev.
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
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.
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
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
A LoRA plus 24-slot decision head over Qwen3.8-27B from Ming Xu's train-jev project: it serves TypeSafe's System One API with 0.851 Kev transfer-v4 accuracy and ECE 0.053, and needs an 80 GB GPU.
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.