Jev directory
CommunityJev Models & Reimplementations
mini-Jev — huggingface.co
An early step-10,626 checkpoint of a 0.6B agentic decision model on Qwen3-0.6B that scores supplied tool and action candidates, reaching 96.5% on a BFCL V1 function-selection adaptation under Apache-2.0.
- Reading time
- 2 min
- Published by
- Community
- Author
- samatv256
- Last verified
- Oct 2026
This is one of several Jev listings for the same project. The canonical entry is mini-Jev — r-ms.
Used inJev for Software & DevTools
- Languages
- Python
- Capabilities
- Open sourceSelf-hostedRuns locally
Highlights
- Give it a state, a question and candidate actions; a permutation-equivariant head returns a probability per candidate.
- BFCL V1 multiple-function adaptation 96.5% (193/200), selecting from 2-4 definitions; not an official leaderboard score.
- Full LocalLLaMA/typed-decisions test 34.0% (680/2000), with Choice 27.0%, Noul 52.2% and Score 25.6%.
- Early checkpoint at 10,626 of a planned 50,000 updates; adapter/ and decision_head.safetensors are the release files.
- CUDA GPU with BF16 required for the 4-bit package; input branches are limited to 8,192 tokens.
Watch out
Apache-2.0 release with a separately licensed Qwen3-0.6B base. This is an interim 10,626-update checkpoint of a planned 50,000; its four evaluation figures come from different datasets and denominators and are not comparable, a 17+ candidate slice scored 4/17, and probabilities are not calibrated for arbitrary domains.
Questions about this entry
Short answers drawn from the linked project or page.
Finite-choice tool and action selection. Given a state, a question and candidate actions, it returns a probability for each candidate through a permutation-equivariant scoring path rather than generating text or tool arguments, normalizing probabilities over the finite set you supply.
They use different datasets and denominators and are not comparable: 96.5% (193/200) on a BFCL V1 function-selection adaptation, 65.6% (63/96) on a bucket-balanced held-out choice slice, 72.0% (36/50) on a synthetic OOD sanity check, and 34.0% (680/2000) on the full LocalLLaMA/typed-decisions test.
No. Root model.safetensors and odm_mini.py are the previous v1 baseline kept for history; the step-10,626 checkpoint loads from adapter/, decision_head.safetensors, inference.py and config.json, and MiniJev.load("samatv256/mini-Jev", revision="step-010626") pins the release.
A CUDA GPU with BF16 support for the 4-bit package, plus the separately downloaded Qwen/Qwen3-0.6B base model. Input branches are limited to 8,192 tokens, and option IDs are bookkeeping excluded from the model text.
Poorly. The card says the model is less reliable for large candidate sets, and a selected 17+ candidate held-out slice scored 4/17, so large sets need particular care; the card also notes its probabilities have not been calibrated for arbitrary domains.
Apache-2.0 for the release, with the Qwen3-0.6B base separately licensed by Qwen under Apache-2.0. Decision supervision draws in part on the CC BY 4.0 Jev Decisions v1 dataset, whose upstream sources carry their own attribution and licence terms.
No. It is an early 10,626-update checkpoint with training planned through 50,000 updates; the card says more training and broader decision supervision are coming, and general typed decisions remain a weakness at 34.0% on the full typed-decisions test.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Standard One SH (StandardThinking)
LinkedIn: 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 NewsHacker News: Pointer to a DiffusionGemma plus vLLM tutorial for a Jev-shaped classifier.
1 pointHow to build a Jev-style classifier with DiffusionGemma and vLLMRead the thread on Hacker NewsMore Jev resources like this
chenz53's standalone decision model derived from Google Gemma 4 26B A4B IT: full merged BF16 weights, a fitted temperature of 1.736507, and 79.50% equal-type accuracy over 1,100 final questions.
Jev Models & Reimplementations#community#open-models#local
Community
An open Thai + English System One decision model at 0.8B, Apache-2.0: choice with up to 255 options, fractional score, and noul answers in one forward pass over a Qwen3.5-0.8B text tower.
Jev Models & Reimplementations#community#open-models#local
Community
Strands Agents' small decider for agent plumbing: routing, tool selection, triage, and guardrails from a 2B LoRA checkpoint that serves /v1/systemone locally or through a Python package.
Jev Models & Reimplementations#community#open-models#local
Community
A pointer-network decision model from Keyush Nisar that reads one agent state document plus typed questions and returns calibrated tool-routing probabilities in a single forward pass, reporting 89.00% zero-shot tool routing.
Jev Models & Reimplementations#community#open-models#local
Community
A DistilBERT prototype from akumar0205 that sorts historical email into legitimate and spam, reporting 0.987 accuracy and 0.983 spam F1 on a held-out SpamAssassin split; it is not a phishing or malware detector.
Jev Models & Reimplementations#community#open-models#local
Community
A small distilled sibling of Perplexity's pplx-decider-v1.2: a Qwen3.5-4B decider that reads a state plus typed choice, noul and score questions in one pass, keeping most of the 27B model's Decision Index.
Jev Models & Reimplementations#community#open-models#multimodal
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.