Jev directory
CommunityJev Models & Reimplementations
W1-JEV
A 3.74B language-diffusion decision model that answers 184 of 231 matching adaptation questions (79.65%) against Jev 1.13.0's public 86.58%, matching it on the easy subset at 33 ms p50 on one RTX 5090.
- Reading time
- 2 min
- Published by
- Community
- Author
- WhaletechAI
- Last verified
- Oct 2026
- Languages
- Python
- Capabilities
- Open sourceSelf-hostedRuns locally
Highlights
- Custom LangDiT language diffusion transformer with 3,739,097,600 parameters, 48 blocks and a 64,512-token vocabulary.
- Answers 184 of 231 matching adaptation questions (79.65%) against Jev 1.13.0's public 200/231 (86.58%).
- Easy subset 48/48; the full-set gap is 6.93 percentage points, and the quoted 92% is a relative ratio.
- 33.13 ms p50 and 374.02 ms p95 on one RTX 5090, 21.32 requests/s and 100% strict format validity across 231 requests.
- Reported GPU occupancy cost $0.00567 per 1,000 requests at $0.435/hour; peak allocated VRAM 7.45 GiB.
Watch out
Publisher-reported numbers: the 92% is a ratio of accuracies on 231 matching questions, not absolute accuracy, and the cost is allocated GPU occupancy excluding service overhead, not a hosted API price. Peak VRAM is an allocator measurement. The repository ships LICENSE, LICENSE.djev and NOTICE files, but the card does not name the licence terms.
Questions about this entry
Short answers drawn from the linked project or page.
A 4B-class language diffusion model for structured decision tasks built on the creators' W1 DLM, with a custom LangDiT transformer, bidirectional attention and timestep conditioning. Post-training aligned its output format to reproduce JEV-style decision tasks.
It is a ratio of accuracies on 231 matching public adaptation questions: W1-JEV's 184/231 (79.65%) against Jev 1.13.0's public 200/231 (86.58%). Absolute accuracy is 79.65%, a 6.93-point gap, and both sides match at 48/48 on the easy subset.
One RTX 5090 with two concurrent clients, one GPU worker and maximum batch size 2 across 116 batches: 33.13 ms p50 and 374.02 ms p95 latency, 21.32 requests/s, 10.8373 s wall clock and 100% strict format validity on all 231 completed requests.
The card reports $0.00567 per 1,000 requests of allocated GPU occupancy at $0.435 per GPU-hour, derived from the run duration of 10.8373 seconds. That figure excludes other service overhead and is not a hosted API price.
Peak PyTorch allocated VRAM was 7.45 GiB in the reported run, an allocator measurement rather than total deployment VRAM. The model holds 3,739,097,600 parameters, ships FP32 in w1-jev.pt and runs BF16 by default.
Use the W1-specific runtime: structured_server.py runs the local HTTP service, examples/ask.py is an executable choice-question client, and config.json holds architecture and context settings. The default 4,096-token input limit is configurable via model.max_seq_len.
The repository includes LICENSE, LICENSE.djev and NOTICE files, but the card does not name the licence terms, so check those files before redistribution or commercial use. The checkpoint carries no training state such as optimizer or EMA weights.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Standard One SH (StandardThinking)
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 NewsMore Jev resources like this
A CPU-first decision model from Decision Models that answers the seven questions an LLM auto-router needs (category, difficulty, stakes, four flags) in one ONNX pass; MIT weights plus a Jev-compatible hosted API.
Jev Models & Reimplementations#community#open-models#local
Community
The v3 Jev-Style decision model at 2B in GGUF: F16 3.78 GB and Q8_0 2.01 GB match PyTorch FP32 on 100% and 99.7% of parity rows, 73.6% on JevBench v1.4.1 public items.
Jev Models & Reimplementations#community#open-models#local
Community
TokenRhythm's 4B prefill-only decision model built on NeoHorse-1-4B: given a state and application-defined questions it returns Choice, Noul, and Score probabilities without generating text, reporting a 77.70 six-group aggregate under Apache-2.0.
Jev Models & Reimplementations#community#open-models#local
Community
A 70.8M-parameter DeBERTa-v3-xsmall cross-encoder that answers typed decision questions in one batched pass with zero generated tokens, shipped with a calibration.json that the engine requires.
Jev Models & Reimplementations#community#open-models#local
Community
A decision model that scores its own clarifying questions by value of information: at AUC .799 against an oracle .797, it asks only when asking is expected to change the answer, then decides or hands off.
Jev Models & Reimplementations#community#open-models#local
Community
An open distillation of Jev 1.13 onto Qwen3.8-27B: a 108.9M-parameter decision block answers typed questions in one prefill pass and adds a POST /v1/decide route beside a full System 2.
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.