Jev directory
CommunityJev Models & Reimplementations
Vega 0.8B: public intents
nandakishorm's Apache-2.0 Vega decision models: a 0.8B default and a 4B under 4b/, both wrapping a frozen Qwen3.5 backbone with a small trained engine for choice, boolean and score questions over a 73,728-token context.
- Reading time
- 3 min
- Published by
- Community
- Author
- nandakishorm
- Last verified
- Oct 2026
Highlights
- 0.8B and 4B sizes sharing one code path; the language model is frozen and never trained, with everything learned in a 57 MB or 124 MB engine.
- 73,728-token context with read-once prefix caching, a memory reader over 16 equal pieces, and refusal instead of silent truncation.
- 0.267 s per decision on a T4 in fp16 and 0.28 s on Apple silicon; the 0.8B answers in 74 ms p50 on the public menu against 162 ms for the hosted model.
- On all 1,075 DecisionBench rows the 4B records calibration error 0.098, NLL 1.420 and KL 0.854 against the hosted model's 0.122, 3.196 and 6.099.
- Images go through the frozen backbone's vision encoder via vega_vision.py, requiring the vision extra and vegaml 0.8.0 or later.
Watch out
Apache-2.0. The card frames its task-level wins as thin (eight of twelve are three rows or fewer and two tasks fail to beat a majority-class baseline) while asking readers to trust the calibration block. The vision path is functional but not covered by the published figures, and vegaml 0.7.1 and earlier cannot reproduce them from PyPI.
Questions about this entry
Short answers drawn from the linked project or page.
An Apache-2.0 decision-model repo from nandakishorm with a 0.8B default at the root and a 4B under 4b/. It answers choice, boolean and score questions, and the card keeps Jev-comparable benchmark tables for both sizes.
import vegaml; vegaml.load() returns the 0.8B default and vegaml.load("4b") the 4B, both on one code path. The 0.8B stays the default because 22 ms vs 28 ms and 57 MB vs 124 MB settle more deployments than four points of accuracy do.
73,728 tokens. A state of at least 4,096 tokens is read once with a KV cache and questions continue from a copy; the frozen backbone carries 262,144 rotary positions, so 73k sits well inside its range.
Partly. The hosted model is more accurate overall, 0.866 vs 0.662 on JevBench and 0.752 vs 0.633 on DecisionBench for the 4B, but Vega wins the calibration block: lower expected calibration error, negative log likelihood and KL on all 1,075 DecisionBench rows.
Yes, using the frozen backbone's vision encoder through vega_vision.py and the vegaml[vision] extra, which carries Pillow and torchvision. It needs vegaml 0.8.0 or later; earlier releases raised ModuleNotFoundError on the first packaged image call, and the published figures are all text.
It is refused, never quietly shortened; the card says a decision made on a truncated contract is worse than no decision. Long states are also cut into 16 equal pieces and mean-pooled so evidence late in a document still reaches the decision.
Apache-2.0, with both sizes in one repository and gated task adapters alongside the trained engines. The headline comparison was run on the 0.8B on 2026-10-11 over two Tesla T4s, 3,841 decisions per system, against the hosted reference called live over the same rows.
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
alibiserikbay's Apache-2.0 Jev alternative on Qwen3.5-4B with a merged rank-16 LoRA: one forward pass returns a probability per option, 0.784 on JevBench's public hard tier at about 13 ms per decision.
Jev Models & Reimplementations#community#open-models#local
Community
The 9B JevK5: the same v0.3 data on Qwen3.5-9B, retrained for 1.5 epochs by a pre-registered sweep, reaching index proxy 0.794 and a 0.859 hand-written hard set, Apache-2.0.
Jev Models & Reimplementations#community#open-models#local
Community
An Apple Silicon MLX bf16 build of the v1 Jev-Style decision model: 77 ms decisions on an M1 Max with calibrated choice, bool and score probabilities, Apache-2.0 weights.
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
A 24 GB-GPU alternative from the train-jev project: Qwen3.5-4B plus a rank-16 LoRA and 24-slot decision head scoring 0.647 PK accuracy, 40.4 Decision Index and 43 ms p50 latency, ahead of Kev-4B at the same size.
Jev Models & Reimplementations#community#open-models#local
Community
A Laya fine-tune for coding decisions: roughly 9k rows from SWE-Gym teach keep/drop file relevance, priority scoring and injection detection, reporting 0.833 accuracy on 1,150 held-out rows at 126 ms p50.
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.