A rank-16 LoRA on Qwen3.5-4B built for the JevBench setting. A document, a policy and a question go in; a calibrated distribution over the option letters comes out of one forward pass. Research and demo use only.
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, classify
choice, score, noul, classify, route
Architecture
hopper
winnow
Fine-tuned from
qwen/qwen3.5-4b
google/gemma-4-12b-it
License
Research and demo use only (training data includes RACE, non-commercial); serving code Apache-2.0
apache-2.0
Availability
Open weights
Open weights
Hosted by
—
—
Input price
—
—
Decision accuracy
68.5%
85.7%
Calibration error
0.102
—
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 hopper and winnow?
hopper is from HopitAI and winnow from EldanRing. Both have open weights you can download and run. Both answer choice, score, noul and classify questions. Only winnow answers route. hopper is the smaller model, at 4.0B parameters to 12B. hopper is licensed other; winnow, apache-2.0.
Which is more accurate, hopper or winnow?
They report on different suites — hopper 68.5% on JevBench public hard tier (111 items, measured by the authors), winnow 85.7% on JevBench public subset (231 items), Q8_0 — so the numbers do not rank them. Test both on your own labelled examples.
Which is cheaper, hopper or winnow?
hopper: Free (open weights). winnow: Free (open weights). Open weights cost nothing per call beyond your own hardware.
Can I run hopper or winnow locally?
Yes, both: systemone pull hopit-ai/hopper and systemone pull eldanring/winnow download the weights.
JevBench public hard tier (111 items, measured by the authors)
