A typed-decision layer for Google's DiffusionGemma, from David Villalón at Maisa AI. It compiles a request into a small answer canvas, runs one denoising read on patched vLLM and reads the probabilities of the allowed labels, for text, images and images offered as options.
AutoTrust's student of TypeSafe Jev 1.13. A frozen Qwen3.8-27B plus a 108.9M-parameter decision block trained on Jev's own output distributions; one set of weights answers typed questions in one pass (System 1) or generates text with the untouched base (System 2).
Decides
choice, score, noul
choice, score, noul
Architecture
djev
blocks-of-experts
Fine-tuned from
google/diffusiongemma-26b-a4b-it
qwen/qwen3.8-27b
License
apache-2.0
apache-2.0
Availability
Open weights + hosted API
Open weights
Hosted by
Maisa
—
Input price
$0.035/MTok
—
Decision accuracy
—
88.7%
Calibration error
—
—
Valid action rate
—
—
Median latency
—
137 ms
p95 latency
—
—
Figures are from each model’s manifest; accuracy and latency are what the publishers report, on their own suites and hardware. Add a third model.
Questions
What is the difference between djev and jev-27b?
djev is from Maisa and jev-27b from AutoTrust AI Lab. djev has open weights and a hosted API; jev-27b has open weights you can download and run. Both answer choice, score and noul questions. djev is the smaller model, at 26B parameters to 27B.
Which is more accurate, djev or jev-27b?
Only jev-27b publishes an accuracy figure (88.7% on JevBench public set (231 items), family-macro score, maker's run); djev does not, so there is no comparison to make without your own test.
Which is cheaper, djev or jev-27b?
djev: $0.035 / $0 per 1M, or free to self-host. jev-27b: Free (open weights). Open weights cost nothing per call beyond your own hardware.
Can I run djev or jev-27b locally?
Yes, both: systemone pull maisa/djev and systemone pull autotrust-ai/jev-27b download the weights.
Evaluation suite
—
JevBench public set (231 items), family-macro score, maker's run
