60 models · checked 20 September 2026
The decision model directory
Every model we can verify exists, with its base, its licence, and a straight answer about whether anyone has measured it. One of them has been. We say so on every card.
How to read this
Decision models went from one product to a whole shelf in about five days. Jev shipped on 15 September 2026; every open-weight model below was published within the week. Nothing that recent has been checked by anyone outside the people who trained it, which is the entire reason this directory grades evidence instead of ranking models.
Benchmarked means we ran it ourselves on a frozen test split and published the numbers with confidence intervals. Community means the facts below come from a public model card and we have measured nothing. There is no middle grade, and we would rather leave a field blank than fill it from a press release.
Browse by
Benchmarked
We ran it, on a frozen split, and the numbers are on the page.
Open weights, not yet benchmarked
Every open-weight model we could verify, including the quantised builds and adapters that other directories drop. Ordered by how much attention each one is getting, which is the only signal available until we have run them. It is not a ranking of quality.
laya
convaiinnovations
The most-liked open-weight decision model of the post-Jev wave, a 421M ModernBERT-large encoder. The base release of a three-model family.
421M · apache-2.0
Community cardQwen-2.5-1B-RLCD
harshatheg
Qwen2.5-1.5B-Instruct fine-tuned to return typed decisions.
Qwen/Qwen2.5-1.5B-Instruct · apache-2.0
Community cardcua-s1-forms
cua-ai
A System One model narrowed to one job, filling forms. MIT licensed, and the clearest example in the directory of the interface spreading past the benchmark it was born on.
mit
Community cardlaya-multilingual
convaiinnovations
The only multilingual decision model we have found, a 322M encoder on mmBERT. Also the one our English-only test set cannot say anything useful about.
322M · apache-2.0
Community carddecider-2b
Mapika
The most-downloaded open-weight decision model in the post-Jev wave. A Qwen3.5-2B decoder fine-tuned to emit typed decisions, and the largest member of a three-model family.
1.9B · Qwen/Qwen3.5-2B-Base · apache-2.0
Community cardlaya-typed-decisions
convaiinnovations
The Laya variant tuned on the typed-decisions benchmark, and the only open model whose card publishes head-to-head numbers against Jev. One of those numbers is the reason we started measuring.
421M · apache-2.0
Community cardopen-jev-deberta-v3-large
com-kotobalabs
A DeBERTa-v3-large encoder positioned as an open reimplementation of Jev. The highest ratio of likes to downloads in the directory, which is usually what a credible claim looks like early.
434M · microsoft/deberta-v3-large · apache-2.0
Community cardLFM2.5-350M-RLCD
notnotsamuel
A typed-decision model from notnotsamuel, 354M parameters. Licensed other, check before commercial use.
354M · other
Community cardrlcd-modernbert-151m
heman10x
The smallest decision model we track, at 151M parameters on a GLiClass/ModernBERT encoder. If it holds accuracy, it changes what this class of model costs to run.
151M · knowledgator/gliclass-modern-base-v2.0 · apache-2.0
Community cardmodernbert-ja-310m-jev
argos1111
modernbert-ja-310m fine-tuned to return typed decisions, 315M parameters. Licensed cc-by-sa-4.0, check before commercial use.
315M · sbintuitions/modernbert-ja-310m · cc-by-sa-4.0
Community cardJEV-CPU
Meanblock
Qwen3-0.6B fine-tuned to return typed decisions.
Qwen/Qwen3-0.6B · mit
Community cardlaya-mlx
aac6fef
laya fine-tuned to return typed decisions, 421M parameters.
421M · convaiinnovations/laya · apache-2.0
Community cardlaya-vision-smolvlm-256m
thaitea
SmolVLM-256M-Instruct fine-tuned to return typed decisions, 237M parameters. Licensed cc-by-nc-sa-4.0, check before commercial use.
237M · HuggingFaceTB/SmolVLM-256M-Instruct · cc-by-nc-sa-4.0
Community cardsystem-one-qwen3.5-4b-scorer
pngwn
The largest open model in the directory and the only one aimed at a single primitive, score. Licensed cc-by-nc-4.0, which rules it out of commercial use regardless of how it benchmarks.
Qwen/Qwen3.5-4B-Base · cc-by-nc-4.0
Community carddecider-2b-vision
Mapika
Qwen3.5-2B-Base fine-tuned to return typed decisions, 2.2B parameters.
2.2B · Qwen/Qwen3.5-2B-Base · apache-2.0
Community cardlaya-onnx
Mattepiu
A ONNX build of laya, for running typed decisions on runtimes without PyTorch instead of a hosted API.
convaiinnovations/laya · apache-2.0
Community cardJeff-1
GestaltLabs
A LoRA adapter that turns Qwen3-4B-Instruct-2507 into a typed-decision model without shipping full weights.
Qwen/Qwen3-4B-Instruct-2507 · apache-2.0
Community cardLFM2.5-2.6B-RLCD
monotykamary
LFM2.5-2.6B fine-tuned to return typed decisions, 2.7B parameters. Licensed other, check before commercial use.
2.7B · LiquidAI/LFM2.5-2.6B · other
Community cardjev-schema-scorer-deberta-v3-large
mobarmg
deberta-v3-large fine-tuned to return typed decisions, 435M parameters.
435M · microsoft/deberta-v3-large · mit
Community cardOpenThai-SystemOne
iapp
Qwen3.5-0.8B-Base fine-tuned to return typed decisions, 753M parameters.
753M · Qwen/Qwen3.5-0.8B-Base · apache-2.0
Community cardsystem-one-qwen3.5-4b-scorer-v2b
pngwn
A LoRA adapter that turns Qwen3.5-4B-Base into a typed-decision model without shipping full weights. Licensed cc-by-nc-4.0, check before commercial use.
Qwen/Qwen3.5-4B-Base · cc-by-nc-4.0
Community cardMobiMind-Decider-7B
IPADS-SAI
A typed-decision model from IPADS-SAI, 8.3B parameters.
8.3B · apache-2.0
Community cardlaya-multilingual-mlx
aac6fef
laya-multilingual fine-tuned to return typed decisions, 322M parameters.
322M · convaiinnovations/laya-multilingual · apache-2.0
Community cardQwen-2.5-1B-RLCD
shreyansh26
Qwen2.5-1.5B-Instruct fine-tuned to return typed decisions.
Qwen/Qwen2.5-1.5B-Instruct · apache-2.0
Community carddecider-2b-GGUF
cosetoenor
A GGUF build of decider-2b, for running typed decisions on llama.cpp instead of a hosted API.
Mapika/decider-2b · apache-2.0
Community cardlaya-mind2web-browser-agent
ShaunSpark
laya fine-tuned to return typed decisions.
convaiinnovations/laya · apache-2.0
Community carddecider-2b-vision-GGUF
mradermacher
A GGUF build of decider-2b-vision, for running typed decisions on llama.cpp instead of a hosted API.
Mapika/decider-2b-vision · apache-2.0
Community cardqwen3-0.6b-rlcd-decision
anthonym21
Qwen3-0.6B-Base fine-tuned to return typed decisions, 596M parameters.
596M · Qwen/Qwen3-0.6B-Base · apache-2.0
Community cardjev-lite
vagmi
A LoRA adapter that turns gemma-4-E4B-it into a typed-decision model without shipping full weights. Licensed gemma, check before commercial use.
google/gemma-4-E4B-it · gemma
Community cardlaya-multilingual-coreml
aac6fef
A Core ML build of laya-multilingual, for running typed decisions on on-device Apple runtimes instead of a hosted API.
convaiinnovations/laya-multilingual · apache-2.0
Community cardlaya-GGUF
mys
A GGUF build of laya, for running typed decisions on llama.cpp instead of a hosted API.
convaiinnovations/laya · apache-2.0
Community cardlaya-multilingual-GGUF
mys
A GGUF build of laya-multilingual, for running typed decisions on llama.cpp instead of a hosted API.
convaiinnovations/laya-multilingual · apache-2.0
Community cardlaya-onnx-fp16
sevenreasons
A ONNX build of laya, for running typed decisions on runtimes without PyTorch instead of a hosted API.
convaiinnovations/laya · apache-2.0
Community carddecider-0.8b-GGUF
mradermacher
A GGUF build of decider-0.8b, for running typed decisions on llama.cpp instead of a hosted API.
Mapika/decider-0.8b · apache-2.0
Community carddecider-0.8b
Mapika
Qwen3.5-0.8B-Base fine-tuned to return typed decisions, 752M parameters.
752M · Qwen/Qwen3.5-0.8B-Base · apache-2.0
Community cardgemma-e2b-rlcd
larkooo
A MLX build of gemma-4-E2B-it, for running typed decisions on Apple Silicon instead of a hosted API.
5.1B · google/gemma-4-E2B-it · apache-2.0
Community cardQwen3-14B-RLCD-Decision-LoRA
Foodoo1
A LoRA adapter that turns Qwen3-14B into a typed-decision model without shipping full weights.
Qwen/Qwen3-14B · apache-2.0
Community cardlaya-coreml
aac6fef
A Core ML build of laya, for running typed decisions on on-device Apple runtimes instead of a hosted API.
convaiinnovations/laya · apache-2.0
Community cardlaya-grounded
Luni
laya fine-tuned to return typed decisions, 421M parameters. Licensed cc-by-nc-4.0, check before commercial use.
421M · convaiinnovations/laya · cc-by-nc-4.0
Community cardlaya-typed-decisions-GGUF
mys
A GGUF build of laya-typed-decisions, for running typed decisions on llama.cpp instead of a hosted API.
convaiinnovations/laya-typed-decisions · apache-2.0
Community cardqwen3-0.6b-rlcd
thefloydd
Qwen3-0.6B-Base fine-tuned to return typed decisions.
Qwen/Qwen3-0.6B-Base · apache-2.0
Community cardMobiMind-Decider-7B-GGUF
mradermacher
A GGUF build of MobiMind-Decider-7B, for running typed decisions on llama.cpp instead of a hosted API.
IPADS-SAI/MobiMind-Decider-7B · apache-2.0
Community cardsystemone-lite-0.5b
dwidlee
A 0.5B decision model on Qwen2.5, the only entry here built on a previous-generation base. Real downloads, zero likes.
494M · Qwen/Qwen2.5-0.5B-Instruct · apache-2.0
Community cardsystem-one-mini
DavidHatley
distilbert-base-uncased fine-tuned to return typed decisions, 69M parameters.
69M · distilbert/distilbert-base-uncased · apache-2.0
Community cardplek-1
Drenel
A typed-decision model from Drenel, 783M parameters.
783M · apache-2.0
Community cardbert4jev
ukung
deberta-v3-large fine-tuned to return typed decisions, 434M parameters.
434M · microsoft/deberta-v3-large · apache-2.0
Community cardJevForge-0.8B
AndeyTait
Qwen3.5-0.8B fine-tuned to return typed decisions.
Qwen/Qwen3.5-0.8B · apache-2.0
Community cardLlama-3B-RLCD-SFT_Llama-3B-Flash_936_full_model
TaylorAI
A typed-decision model from TaylorAI.
Hugging Face
Community cardsystem-one-distilled
shreyanbr
deberta-v3-xsmall-zeroshot-v1.1-all-33 fine-tuned to return typed decisions, 71M parameters.
71M · MoritzLaurer/deberta-v3-xsmall-zeroshot-v1.1-all-33 · apache-2.0
Community cardjev-0.5b
jaswanthsanjay88
A LoRA adapter that turns Qwen2.5-0.5B into a typed-decision model without shipping full weights.
Qwen/Qwen2.5-0.5B
Community cardsystem-one-gold
shreyanbr
deberta-v3-xsmall-zeroshot-v1.1-all-33 fine-tuned to return typed decisions, 71M parameters.
71M · MoritzLaurer/deberta-v3-xsmall-zeroshot-v1.1-all-33 · apache-2.0
Community cardsystem-one-zeroshot
shreyanbr
deberta-v3-xsmall-zeroshot-v1.1-all-33 fine-tuned to return typed decisions, 71M parameters.
71M · MoritzLaurer/deberta-v3-xsmall-zeroshot-v1.1-all-33 · apache-2.0
Community cardlaya-multilingual-coreml-ane
aac6fef
A Core ML build of laya-multilingual, for running typed decisions on on-device Apple runtimes instead of a hosted API.
convaiinnovations/laya-multilingual · apache-2.0
Community cardlaya-multilingual-coreml-ane-w8
aac6fef
A Core ML build of laya-multilingual, for running typed decisions on on-device Apple runtimes instead of a hosted API.
convaiinnovations/laya-multilingual · apache-2.0
Community cardlaya-typed-decisions-coreml
aac6fef
A Core ML build of laya-typed-decisions, for running typed decisions on on-device Apple runtimes instead of a hosted API.
convaiinnovations/laya-typed-decisions · apache-2.0
Community cardlaya-multilingual-coreml-snake
aac6fef
A Core ML build of laya-multilingual, for running typed decisions on on-device Apple runtimes instead of a hosted API.
convaiinnovations/laya-multilingual · apache-2.0
Community carddecider-35b-a3b
Mapika
Qwen3.5-35B-A3B-Base fine-tuned to return typed decisions, 35B parameters.
35B · Qwen/Qwen3.5-35B-A3B-Base · apache-2.0
Community carddecider-35b-a3b-nvfp4
Mapika
A typed-decision model from Mapika, 18B parameters.
18B · Mapika/decider-35b-a3b · apache-2.0
Community cardlaya-typed-decisions-mlx
aac6fef
laya-typed-decisions fine-tuned to return typed decisions, 421M parameters.
421M · convaiinnovations/laya-typed-decisions · apache-2.0
What we left out
Searching Hugging Face for this category turns up far more repositories than appear here. We dropped empty repositories with no weights, downloads or likes; exact copies of another account’s weights; and projects that only matched on a name, like a 3D scene model whose owner happens to be called Jev. Same-author variants trained differently were kept, because those are genuinely different models.
Everything that remains comes from each repository’s own metadata. A directory padded with entries nobody checked is worth less than a short one.
None of these were measured on your data
Public benchmarks tell you which models are worth a shortlist. They cannot tell you which one clears your error budget. That takes your own historical decisions, frozen into a test set.