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.
A Gemma 4 12B fine-tune for typed decisions that also chats and reads images. Ships as GGUF for llama.cpp, holds a 64K context on a 16 GB GPU, and serves /v1/systemone next to /v1/chat/completions.
Decides
choice, score, noul
choice, score, noul, classify, route
Architecture
djev
winnow
Fine-tuned from
google/diffusiongemma-26b-a4b-it
google/gemma-4-12b-it
License
apache-2.0
apache-2.0
Availability
Open weights + hosted API
Open weights
Hosted by
Maisa
—
Input price
$0.035/MTok
—
Decision accuracy
—
85.7%
Calibration error
—
—
Valid action rate
—
—
Median latency
—
143 ms
p95 latency
—
—
Evaluation suite
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 winnow?
djev is from Maisa and winnow from EldanRing. djev has open weights and a hosted API; winnow has open weights you can download and run. Both answer choice, score and noul questions. Only winnow answers classify and route. winnow is the smaller model, at 12B parameters to 26B.
Which is more accurate, djev or winnow?
Only winnow publishes an accuracy figure (85.7% on JevBench public subset (231 items), Q8_0); djev does not, so there is no comparison to make without your own test.
Which is cheaper, djev or winnow?
djev: $0.035 / $0 per 1M, or free to self-host. winnow: Free (open weights). Open weights cost nothing per call beyond your own hardware.
Can I run djev or winnow locally?
Yes, both: systemone pull maisa/djev and systemone pull eldanring/winnow download the weights.
