Jev directory
CommunityJev Models & Reimplementations
Robo-Jev 2B
Steerix's robo-jev v0.7 merges a robotics LoRA into Qwen3.5-2B as a 2B System 1 that answers physical-judgment questions from a robot's scene; the card reports 0.81 physical judgment against TypeSafe Jev's 0.28.
- Reading time
- 2 min
- Published by
- Community
- Author
- Steerix Org
- Last verified
- Oct 2026
- Languages
- Python
- Capabilities
- Open sourceSelf-hostedRuns locally
Highlights
- Merged weights: Qwen3.5-2B with the robo-jev v0.7 LoRA merged, saved in bfloat16 as a text-only Qwen3_5ForCausalLM.
- Self-reported physical judgment of 0.81 versus TypeSafe Jev 0.28, Liquid d1 0.17 and untrained Qwen3.5-2B minus 0.15.
- General reading (BoolQ + MultiNLI) is 0.71, below Jev at 0.91 and Liquid d1 at 0.93, a specialization trade-off.
- One forward pass per question: about 50 ms for a short request on one L4 GPU, a dozen questions in about 0.1 s.
- Supports 1 to 12 questions per request, 2 to 466 choice options and state text up to 16,000 characters.
Quickstart
bash
pip install torch "transformers>=5.17"
hf download Steerix-org/robo-jev-2b robo_jev_decide.py --local-dir .
python robo_jev_decide.py --model Steerix-org/robo-jev-2b --request request.jsonWatch out
Card reviewed 2026-10-17; the 0.81 physical-judgment score is self-reported on robo-jev's own benchmark of 1,100 held-out situations, and comparison models were asked through the OpenRouter Decisions API while robo-jev ran locally. The card also notes options past the 26th get two-letter codes that were not part of training, so transfer beyond the authors' benchmark is unproven.
Questions about this entry
Short answers drawn from the linked project or page.
A fast, single-pass 2B System 1 for robots' physical judgment. It reads a short text of what a robot senses and answers questions like whether the glass will tip if it brakes now, with a probability for every answer.
It uses the Decisions API's request shape: a state text plus up to 12 typed questions and a calibrated probability for every answer, from one forward pass per question with no sampling and no reasoning tokens.
The merged weights: Qwen3.5-2B with the robo-jev v0.7 LoRA adapter merged in, saved in bfloat16 as a text-only Qwen3_5ForCausalLM. The base model's vision tower is left out; the adapter alone is at Steerix-org/robo-jev-2b-lora.
On one L4 GPU a short request answers in about 50 ms and a dozen questions in about 0.1 s, because there is one forward pass per question and no token generation at all.
Physical judgment 0.81 versus TypeSafe Jev 0.28, Liquid d1 0.17 and untrained Qwen3.5-2B minus 0.15 on 1,100 held-out situations, but general reading (BoolQ + MultiNLI) is 0.71 against Jev's 0.91: a specialization trade-off.
Steerix built it as a LoRA fine-tune of Qwen/Qwen3.5-2B, released under Apache-2.0 as v0.7 merged weights; the card compares other decision models asked through the OpenRouter Decisions API.
choice questions take 2 to 466 options, 1 to 12 questions per request, and state text up to 16,000 characters; options past the 26th get two-letter codes that were not part of training, and the card's text cuts off while describing that limit.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Standard One SH (StandardThinking)
Hacker News: Pick Verify: a single decision at the warehouse pick station catches confusable picks.
Show HN: Verifying robot picks with a decision modelRead the thread on Hacker NewsLinkedIn: What multimodal decision models actually change, and what they do not.
LinkedInJevDirectory.org
Hacker News: Decision Index 0.2.1 scores open reproductions of Jev on a frozen suite.
jackbravo4 pointsDecision Index 0.2.1 – 70 open reproductions of TypeSafe Jev's Decision ModelRead the thread on Hacker NewsMore Jev resources like this
Multimodal finite-choice decision interface and MuJoCo robotic-arm workbench: dual-camera images and state go to a self-hosted Qwen3.5-9B with a corrective head, which returns one typed next-skill choice without generating text.
Jev Models & Reimplementations#community#robotics#multimodal
Community
shapsider/OmniJevPython
shapsider/OmniJevPythonA LoRA plus 24-slot decision head over Qwen3.8-27B from Ming Xu's train-jev project: it serves TypeSafe's System One API with 0.851 Kev transfer-v4 accuracy and ECE 0.053, and needs an 80 GB GPU.
Jev Models & Reimplementations#community#open-models#local
Community
A 24 GB-GPU alternative from the train-jev project: Qwen3.5-4B plus a rank-16 LoRA and 24-slot decision head scoring 0.647 PK accuracy, 40.4 Decision Index and 43 ms p50 latency, ahead of Kev-4B at the same size.
Jev Models & Reimplementations#community#open-models#local
Community
A Laya fine-tune for coding decisions: roughly 9k rows from SWE-Gym teach keep/drop file relevance, priority scoring and injection detection, reporting 0.833 accuracy on 1,150 held-out rows at 126 ms p50.
Jev Models & Reimplementations#community#open-models#local
Community
A Laya fine-tune from Convai Innovations on the 1,200-case typed-decisions benchmark: it reports 0.777 accuracy on the official 400-case test set, above TypeSafe Jev 1.13.0 at 0.727 and the teacher ceiling at 0.735.
Jev Models & Reimplementations#community#open-models#local
Community
A 9B bilingual Turkish and English Jev-class decision model from Berkay Hasret: a 173 MB LoRA adapter on Qwen/Qwen3.5-9B that answers up to 16 typed questions about one text in a single forward pass and reports 0.783 macro accuracy on full JevBench.
Jev Models & Reimplementations#community#open-models#local
Community
From the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.