Dohnuts
Small multimodal System One models that choose, judge or score text, documents and images in one forward pass.
Overview
Dohnuts builds small multimodal models for direct decisions. Give the model a message, a document or an image and ask it to choose, judge or score; it returns probabilities in a single forward pass, with multiple questions sharing the same input. The 0.8B model, a LoRA on a Qwen3.5 base with a saved calibration, runs locally on a consumer GPU.
The Python toolkit covers training (joint RLCD and cross-entropy on a public-data mixture), checkpoint selection, calibration, evaluation and agent integration. According to a list summary, the model card publishes the benchmark it loses and states that its confidence field is not a measured probability of correctness, and compares against Jev and Laya multilingual. Code is Apache-2.0, but model weights are CC BY-NC-SA 4.0, which restricts commercial use.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which team should handle the request? | billing · technical support · sales |
| Yes / no | Is a refund requested? | yes · no |
How it works
- Load the checkpoint with Predictor.from_checkpoint
- Pass a message, document or image with typed questions
- Read per-question distributions from one forward pass
- Optionally retrain and recalibrate with the provided pipeline
Strengths
- Multimodal inputs
- Runs locally on a consumer GPU
- Full training and evaluation pipeline
- Candid model card
Caveats
- Weights are non-commercial (CC BY-NC-SA 4.0)
- Confidence is not a measured correctness probability
Getting started
Predictor.from_checkpoint("PsiACE/Dohnuts-0.1.0-0.8B").predict(state, questions); see docs/inference.md for setup.At a glance
- Type
- Model
- Status
- research
- Language
- Python
- Runtime
- PyTorch
- Has code
- yes
- License
- Apache-2.0
- Stars
- 36
Use cases and integrations
How curated lists describe it
Dohnuts builds small multimodal models for direct decisions. -> System One model
wh000wh000/awesome-jev-liveDohnuts builds small multimodal models for direct decisions. -> System One model.
hellogumbo/awesome-jevSmall multimodal models for direct decisions on text, documents and images. Its model card publishes the benchmark it loses and states that its confidence field is not a measured probability of correctness. Weights are CC BY-NC-SA.
MrJev/awesome-jev
Frequently asked questions
Can I use the weights commercially?
The weights are CC BY-NC-SA 4.0, which does not allow commercial use; the code is Apache-2.0.
Does it handle images?
Yes, add a PIL image to the state.
How does it compare with Jev?
The model card includes comparisons and reports benchmarks where it loses.
Related projects
Free hosted Jev-style multimodal decision model API (about 20B active) that picks actions from images and options, aimed at robot control.
ServicebetaComfyUI nodes for Jev-style noul, choice and score decisions over images and text using a local imajev adapter on Qwen3.5.
ProjectbetaPythonResearch code for an image-and-text System One model that reads option-letter logits from a multimodal LLM, with training and calibration.
ProjectresearchPythonIndependent open model on Molmo2-ER that answers spatial questions by choosing from fixed options with one shared classifier head.
ProjectresearchPythonOpen project training a vision-enabled Jev-like model (0.8B, 9B) that judges dynamic candidate sets over images with Choice, Noul and Score.
ProjectresearchPythonIndependent Jev-style server on DiffusionGemma and SGLang that classifies doodles, flowers and messages with Choice, Score and Noul.
ServiceresearchPython