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.
StartLux's decision family in five sizes, 0.8B to 27B. All questions in a request are answered in one forward pass, with a probability for every option, through a TypeSafe /v1/systemone-compatible server that records CUDA graphs for short requests.
Decides
choice, score, noul
choice, score, noul
Architecture
djev
startlux-decision
Fine-tuned from
google/diffusiongemma-26b-a4b-it
—
License
apache-2.0
cc-by-nc-4.0
Availability
Open weights + hosted API
Open weights
Hosted by
Maisa
—
Input price
$0.035/MTok
—
Decision accuracy
—
88.3%
Calibration error
—
—
Valid action rate
—
—
Median latency
—
26 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 startlux-decision?
djev is from Maisa and startlux-decision from StartLux. djev has open weights and a hosted API; startlux-decision has open weights you can download and run. Both answer choice, score and noul questions. startlux-decision is the smaller model, at 4.7B parameters to 26B. djev is licensed apache-2.0; startlux-decision, cc-by-nc-4.0.
Which is more accurate, djev or startlux-decision?
Only startlux-decision publishes an accuracy figure (88.3% on JevBench public set (231 items; 204 correct), maker's run); djev does not, so there is no comparison to make without your own test.
Which is cheaper, djev or startlux-decision?
djev: $0.035 / $0 per 1M, or free to self-host. startlux-decision: Free (open weights). Open weights cost nothing per call beyond your own hardware.
Can I run djev or startlux-decision locally?
Yes, both: systemone pull maisa/djev and systemone pull startlux/startlux-decision download the weights.
Evaluation suite
—
JevBench public set (231 items; 204 correct), maker's run
