jev-spatial
Independent open model on Molmo2-ER that answers spatial questions by choosing from fixed options with one shared classifier head.
Overview
Jev-Spatial is an independent research model inspired by Jev's System One idea and built on Allen AI's Molmo2-ER. Every spatial task becomes images plus a question plus candidate options, answered by one shared classifier head with no text generation. Classification tasks (spatial relations, directions, yes/no) are a single choice; numeric estimates such as lengths in metres use two rounds of range then sub-range selection; pointing uses three rounds of picking a cell in a 3 by 3 grid, zooming in after each round.
The README reports it performs close to the autoregressive Molmo2-ER baseline overall, within 1 to 4 points on classification (CV-Bench -1.0, SAT -4.0), with lower metric error on VST (0.472 versus 0.520 m), and better pointing (+2.0 on RefSpatial-Bench, +7.0 on Where2Place). With eight questions per image it is about 4.2 times faster than one-at-a-time AR generation. It is explicitly not affiliated with TypeSafe and not trained on Jev outputs; weights are on Hugging Face.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which option answers the spatial question about these images (relation, direction or yes/no)? | — |
| Choice | Which value range, then sub-range, contains the measured length? | — |
| Choice | Which cell of the 3x3 grid contains the target point (repeated with zoom)? | — |
How it works
- Encode images and question with candidate options
- Shared head scores options and returns one choice
- For regression or pointing, refine over two or three rounds
- Output is always a valid option, so nothing needs parsing
Strengths
- No malformed output; every answer is from a fixed set
- README reports about 4.2x faster than one-at-a-time AR on 8 questions per image
- Better pointing accuracy than the AR baseline per the README
- Open weights on Hugging Face
Caveats
- Research model, not affiliated with TypeSafe
- Slightly lower classification accuracy than the base model on some benchmarks
At a glance
- Type
- Project
- Status
- research
- Language
- Python
- Runtime
- PyTorch / Hugging Face
- Has code
- yes
Use cases and integrations
How curated lists describe it
Jev-Spatial is a System One spatial model inspired by Jev. It builds on Molmo2-ER, and instead of generating text it answers spatial questions by choosing from a fixed set of options with one shared classifier head.
hellogumbo/awesome-jev
Frequently asked questions
Is Jev-Spatial related to TypeSafe's Jev?
Only in idea. The README says it is independent, not endorsed by TypeSafe and not trained on Jev outputs.
How does it do pointing without generating coordinates?
It picks a cell in a 3x3 grid three times, zooming into the chosen cell after each of the first two rounds.
How does it compare with the base model?
Close overall, slightly lower on some classification benchmarks, lower metric error and better pointing per the README.
Related projects
Free hosted Jev-style multimodal decision model API (about 20B active) that picks actions from images and options, aimed at robot control.
ServicebetaOmniJev: omni-modal 4B decision model answering typed questions about images, video, screens and robot scenes in one pass.
ModelresearchPythonComfyUI 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.
ProjectresearchPythonRepository serving CLM-8B, a contrastive System One model behind a TypeSafe-compatible API, aimed at humanoid robot decisions.
ProjectresearchPythonOpen project training a vision-enabled Jev-like model (0.8B, 9B) that judges dynamic candidate sets over images with Choice, Noul and Score.
ProjectresearchPython