layaLaya is an Apache-2.0 multilingual non-autoregressive decision model that answers typed questions (choices, scores, booleans) over text or JSON states with calibrated probabilities in a single ~33 ms forward pass, auto-routing across three checkpoints covering 100+ languages.
JEV-27B-VLJEV-27B-VL adds vision to JEV-27B as an adapter on Qwen3.8-27B, outputting calibrated probabilities for typed decisions (yes/no, 2–16 options, 0–5 ratings) over images and text in one forward pass — enabling zero-shot cold-start video recommendation and benchmark-topping agent/multimodal judging.
JEV-35BJEV-35B is a System 1 decision model fine-tuned from Qwen3.5-35B-A3B (35B MoE, ~3B active) that returns calibrated choices among up to 256 options in one ~200 ms pass for text, vision and control tasks (Decision Index 0.3 public: 59.73).
clefClef is a 27B multimodal model post-trained from Qwen3.8-27B that maps a state (text, JSON, images, or video) plus a schema of typed questions (true/false, choice, or score) to per-option probabilities for every question in a single forward pass—no free-form generation or output parsing—via a joint schema head and a Jev/SystemOne-compatible API.
AutoTrust-JEV-27B-v1.1AutoTrust-JEV-27B-v1.1 is a vLLM-served 27B decision model that returns option probabilities for choice and yes/no questions, reaching 67.31 on the 37-benchmark Decision Index public suite, with its strongest results in tool selection, intent detection, and automation decisions.
clef-flashClef-Flash is a 9B multimodal model post-trained from Qwen3.5-9B that maps a state (text, JSON, images, or video) plus a schema of typed questions to per-option probabilities for every question in a single forward pass, with no free-form generation or parsing.
CLM-v0.1-8BCLM-v0.1-8B is a contrastive "System One" model that trains small state and action projection heads over a frozen Qwen3-8B encoder to score and rank candidates for agentic decision-making, with separable state/action embeddings enabling caching for up to 13× faster inference.
openjevopenjev fine-tunes Qwen3.5 into a zero-shot NLI cross-encoder (4B text+image v2, plus a 35B-A3B MoE variant) that scores premise–hypothesis pairs as entailment/contradiction/neutral — one primitive for reranking, answer grading, content guarding, and driving real-time game agents like Doom and Minecraft.
Julia-1Julia-1 is a 144.3M-parameter multilingual decision model built on mmBERT-small that maps a state, a typed question, and 2–20 supplied options to a scored choice, rubric score, or Boolean probability for routing and typed-decision workloads.
JEV-27BAn open dual-mode model serving calibrated typed decisions (yes/no, 2–16-option choice, 0–5 scores) in one forward pass and ordinary Qwen3.8-27B text generation from a single set of weights in one vLLM engine, while reproducing TypeSafe Jev 1.13's output distributions to within ~0.017 KL.
Jev-OmniJev-Omni is an Apache-2.0 multimodal decision classifier built on Gemma 4 12B IT that returns calibrated per-option probabilities for text, image, audio, and video questions, scoring 63.10% on MMAU and 53.10% on MVBench with millisecond-scale inference.
GEV-26B-DecideA LoRA adapter on Gemma-4-26B-A4B-it (≈4B active parameters) that returns calibrated probabilities for typed decisions—yes/no, 2–256 options, 0–5 ratings—over text and images in one ~45 ms pass, optionally escalating uncertain cases to System 2 reasoning for higher accuracy.
d3d3 is a 27B multimodal decision model that answers choice, yes/no, and score questions about text, JSON, images, or videos in a single call, returning a probability for every option without generating text—built on Qwen3.8-27B for the vLLM Semantic Router.
openjevOpenJev is an open-weights decision model that answers user-defined questions—choice, yes/no probability, or score over up to 52 options—in a single forward pass (~80–210 ms on one H100), delivering calibrated, parse-free decisions for text routing, moderation, model judging, and browser/desktop agent control, with multilingual support and quantized builds.
pplx-decider-v1.1-27bApache-2.0 Qwen3.8-27B finetune for scored candidate decisions, shipping a Qwen3_5Model backbone with a separate BF16 [255,5120] readout, noncausal full-attention behavior, and calibrated predict code that raises Decision Index to 61.56 (v1: 56.4, Jev: 57.9).
kev-4bKev-4B is a LoRA adapter plus pointer head on Qwen3.5-4B-Base that maps a document and typed questions to calibrated per-question probability distributions in one forward pass—no generation—serving TypeSafe's /v1/systemone decision contract at 0.837 out-of-domain locked-test accuracy.
JEV-35B-NVFP4JEV-35B-NVFP4 is a System 1 decision model that scores up to 256 options with calibrated probabilities in one pass, with its 10,240 routed experts quantized to NVFP4 so it runs in 23 GiB of GPU memory instead of 66 GiB while keeping 96.6% answer agreement and essentially unchanged Decision Index and vision scores.
laya-typed-decisionsA 421M-parameter non-autoregressive ModernBERT-large model, RLCD-fine-tuned to emit calibrated probability distributions over typed primitives (choice, score, noul) for four workflows—agent observability, customer service, invoicing, security incidents—at 0.766 accuracy.
JEVAn open-weights 9B decision model (Qwen3.5-9B backbone with a LoRA-tuned 24-slot linear head) distilled from TypeSafe Jev 1.13 that returns calibrated probability distributions for yes/no, 2–16-option choice, and 0–5 score questions in a single forward pass.
JEV-9BJEV-9B is an open dual-mode model that adds a 40.2M-parameter decision adapter to a frozen Qwen3.5-9B backbone, returning calibrated probability distributions for typed decisions (yes/no, 2–16 choices, 0–5 scores) in ~90 ms while leaving base generation byte-identical.
clef-omniClef-Omni is a 30B-A3B MoE multimodal model post-trained from Qwen3-Omni that maps a text/JSON/image/audio/video state plus a typed question schema to probabilities for every option of every question in a single forward pass, with no free-form output.
pplx-decider-v1-27bpplx-decider-v1-27b is an Apache-2.0 PyTorch decision model fine-tuned from Qwen3.8-27B that predicts labeled choices or yes/no answers with calibrated probabilities, supports image inputs, and reaches 85.71% overall accuracy across 11 benchmarks.
d3-turbod3-turbo is a 26B multimodal decision model that reads text, JSON, images, and video once per request to answer many questions at once—choices, yes/no, scores, object boxes, and video tracks—returning calibrated probabilities without generating text.
levlev is a LoRA adapter on Qwen3.5-4B that answers typed yes/no, choice, and score questions about a state in one batched forward pass, returning calibrated probabilities over your options with zero output tokens for routing, moderation, triage, and gating decisions.
cua-s1-formscua-s1-forms is a 706K-parameter single-pass option scorer that maps a GUI form element plus typed candidates (fill each extracted entity, check, click, skip) to per-option probabilities, letting cua-driver fill and submit forms without an autoregressive LLM.
decider-2bdecider-2b is a 1.9B-parameter model that returns calibrated probability distributions over explicit option lists for multiple typed questions in one forward pass—no text generation—for fast routing, classification, judgments, and agent decisions.
h2o-lightning-4bH2O-Lightning-4B is a 4B-parameter multimodal decision model (built on Qwen3.5-4B) that answers choice, yes/no, and ordinal-score questions about text records and attached images in one forward pass returning calibrated per-option probabilities, ranking #1 on JevBench's open-weight composite while running on stock vLLM via a lightweight shim at ~$0.021 per 1,000 decisions.
MiniCPM-OCR-JevMiniCPM-OCR Jev is a 2B vision-language model that scores candidate OCR results for text, formula, and table regions from 0 to 1, enabling automatic acceptance of reliable output, human-review routing, and best-of-N selection across OCR models.
Intern-Decision-4BIntern-Decision-4B is a multimodal Qwen3.5-4B fine-tune that scores every field of a typed decision schema (choice, score, or binary, with up to 8 images) in a single calibrated forward pass—returning per-option probability distributions without generation—achieving a 90.02 benchmark average at ~44 ms per query on an RTX 4090.
THX-01THX-01 is a 322M-parameter non-autoregressive multilingual decision model that answers typed natural-language questions—choices, yes/no, ordinal scores, numbers, verbatim excerpts, and citations—about a given document or state in a single ~10 ms pass, returning calibrated probabilities trained via RL on strictly proper scoring rules.
NanoJevNanoJev is a 0.6B Qwen3-based decision model whose shared backbone and attention Choice head score each state/question-encoded candidate set and return action distributions via softmax—without output-token decoding—for Maze, Snake, ViZDoom Basic, and Predict Position through a structured interface.
Winnow-12BWinnow-12B is a GGUF-packaged LoRA fine-tune of Gemma 4 12B IT for local inference, providing typed decisions (noul, choice, score) via shared-state logit reads without generating answer text, plus ordinary chat and vision, in BF16 and Q8_0 variants tested at 64K context on a 16 GB GPU.
open-jev-deberta-v3-largeA DeBERTa-v3-large encoder that maps one input state plus any number of typed questions (choice, ordered score, or yes/no) to calibrated per-question probability distributions in a single forward pass, reaching 0.854 in-domain accuracy at 0.022 ECE with zero structured-output errors.
JEV-CPUJEV-CPU is a CPU-only port of the SemIf semantic-if engine that loads Qwen3-0.6B in float32 and returns typed option probabilities read directly from option-letter logits in a single forward pass—no GPU or text generation—exposed via CLI and a dependency-free web UI.
d3-edged3-edge is a 0.59B multimodal decision model that answers choice, yes/no, and score questions over text, images, and video in one call—returning a probability per answer without generating text—at ~6.6 ms median latency, servable via Transformers or vLLM Semantic Router.
drex-v1.5Drex v1.5 is a 9B-parameter decision model that scores typed questions (choice, yes/no, ordinal score) against a text or JSON state, returning per-option probabilities in one non-generative forward pass per question via the TypeSafe-compatible `/v1/systemone` API served from Python, llama.cpp, or Ollama.
autojev-27bAutoJev-27B is a 27B-parameter multimodal decision model finetuned from Qwen3.8-27B that returns calibrated probabilities for supplied choices in a single forward pass, exposes a TypeSafe-compatible API and playground, and improves benchmark accuracy/ECE/Brier over its base.
drex-max-v1Drex Max is a dense 27B-parameter multimodal decision model that answers typed questions (choice, noul, score) over text, JSON, and image context, returning probabilities for up to 255 options in a single forward pass without generating text.
Open-Jev-9BOpen-Jev-9B is a LoRA adapter plus scalar decision head for Qwen3.5-9B that scores caller-supplied candidates and returns typed choice/yes-no/ordinal decisions with calibrated probabilities—without autoregressive generation—via a pinned-revision inference server.
Quyet-1.0-LargeQuyet-1.0-Large is a 31.3B-parameter Apache-2.0 decision model (a Gemma-4-31B-it LoRA fine-tune, English/Vietnamese) that, given a text, JSON, or conversation state plus typed questions—choice, score, or true/false—returns one answer per question with calibrated probabilities.
Mica-v0.5-4BMica-v0.5-4B is an Apache-2.0 4B parameter classification model finetuned from Qwen3.5-4B that returns calibrated probability distributions over up to 255 allowed answers for decision questions on inputs up to 32k tokens in English, Korean, or Chinese, offering order-insensitive predictions, configurable abstention, and sub-20ms latency on a single consumer GPU for guardrails, RAG checks, routing,
basal-omni-1.0-ultrabasal-omni-1.0-ultra is an experimental 31B Gemma-4 fine-tune that answers typed questions (choice, yes/no, score, tool selection) over Polish/English text states and zero-shot images by returning a calibrated probability per allowed option—up to 255—in one forward pass, enabling automated decisions with confidence-based deferral.
agent-jevAgentJev-0.6B turns a 2,048-token state and typed boolean, choice, or score questions into full probability distributions in one forward pass, giving agent loops fast, thresholdable gates, routes, and rubric scores without decoded text, JSON repair, or a larger writer model.
rlcd-modernbert-151mA 151M-parameter ModernBERT-based decision engine delivering calibrated choices, scores, and probabilities—plus an explicit abstention slot—in a single ~35ms non-autoregressive forward pass, deployable in-browser via WebGPU/WASM or via ONNX for edge inference.
clef-flash-4bitclef-flash-4bit is a 4-bit MLX quantization of Cloudflare's 9B clef-flash for Apple Silicon whose bundled clef_mlx.py loader (CLI or local SystemOne server) maps a text/JSON/image/video state plus a schema of typed questions to option probabilities in a single forward pass.
Tev1-4B-experimentalTev1-4B-experimental is an experimental supervised fine-tune of Qwen3.5-4B that answers structured decision tasks, returning exactly one option letter given a state, question, and 2–24 labeled choices via a standard next-token head under deterministic decoding.
vega-08b-public-intentsA local decision model that layers a small trained physics engine over a frozen multimodal Qwen3.5-0.8B backbone, returning calibrated choice/boolean/score answers with conformal sets and abstain flags over 73,728-token contexts in ~22 ms per decision, rivaling hosted APIs.
K2-Type-0.9BK2-Type-0.9B is a 0.9B non-generative decision model that scores every option of typed yes/no, multiple-choice, and ordinal questions against one shared text/JSON state with calibrated probabilities in a single forward pass (JevBench public 176/231, ~27 ms median).
NeoHorse-Jev-4BA 4B structured decision model using prefill-only inference to output probabilities over application-defined Choice, Noul, and Score answers, enabling routing, tool selection, and rating in agent workflows, with local deployment via vLLM, SGLang, or native runtimes.
span-02Span-02 is an NVFP4/FP8-quantized Gemma 4 26B MoE fine-tune (~4B active params) served by its own Rust/CUDA engine, answering typed questions (choice, yes/no, score, number) about a text/JSON state plus up to 4 images with calibrated probabilities at 33 ms median latency.
APUS-OpenJev-v1APUS-OpenJev-v1 is a family of Qwen3.5-based decision models (4B, 9B, 35B-A3B) that score candidate actions for browser agents and business workflows, offering selectable low/high compute depth, candidate-aware output projection, and English-Chinese support.
Lumma-fev-27BLumma-Fev-27B is a 27B-parameter decision model fully fine-tuned from Qwen3.8-27B that, given a text or JSON state plus typed questions, returns per-option probabilities in a single forward pass in English, Hindi, Tamil, and Kannada, via the lumma-fev package.
drex-dlmDrex DLM is an 8B diffusion-language decision model that answers typed questions—named choices, yes/no, and ordinal scales—about a supplied context in a single forward pass plus pointer head, returning option probabilities for agentic routing and decision tasks.
drex-v2.0Drex v2.0 is a 35B-parameter mixture-of-experts decision model that scores typed questions against a given context via a pointer head, returning calibrated probabilities for every option in a single forward pass without generation, runnable locally through Python, llama.cpp, or Ollama.
laya-GGUFGGML-compiled via ggmlc (not llama.cpp-compatible) F16/Q8_0/Q4_K_M weights of the 421M ModernBERT-based Laya English System-1 decision model, scoring typed choice/score/noul questions in one encoder pass for email routing, injection guards, and agent gating.
Jeff-Qwen3.5-0.8BJeff-Qwen3.5-0.8B is an Apache-2.0 fine-tune of Qwen3.5-0.8B that turns arbitrary states and plain-language options into calibrated per-option probabilities in a single ~22 ms forward pass, enabling fast local zero-shot classification, routing, and game-playing decisions.
MiniCPM-OCR-Jev-ggufGGUF quantizations (F16 down to IQ3_XS) of the 2B MiniCPM-OCR Jev model plus its required mmproj vision projector for llama.cpp, enabling local 0–1 quality scoring of OCR candidates across text, formula, and table regions.
Tev1-0.8B-experimentalTev1-0.8B-experimental is a 0.8B-parameter supervised fine-tune of Qwen3.5-0.8B that, given a structured prompt containing a state, question, and 2–24 labeled options, deterministically returns a single option letter—making it useful as a lightweight decision/classification layer where application code maps letters back to semantic keys.
matilda-jev-v1A 26.1B-parameter decision model that returns temperature-calibrated probabilities over the options of choice, yes/no, and scored-rubric questions in a single deterministic forward pass—no sampling or answer parsing—scoring 59.26 on the Jev Decision Index.
MedJev-Qwen3.5-0.8BMedJev-Qwen3.5-0.8B is a LoRA adapter plus pointer head on Qwen3.5-0.8B-Base that extracts 11 predefined clinical variables from free-text clinical notes in a single forward pass (~62 ms per note), returning calibrated probability distributions over allowed answers without text generation, achieving 0.878 micro accuracy against silver GPT-4-derived labels.
ModernJEV-Decide-PreviewA 149.6M-parameter ModernBERT cross-encoder fine-tuned on 60,000 agent decisions that scores and ranks user-supplied candidate labels, given a conversation, policy, and tool descriptions, for next-action routing and tool selection within 4,096-token inputs.
Open-Jev-2BA LoRA adapter and scalar decision head for Qwen3.5-2B that directly scores caller-supplied candidates and returns calibrated typed decisions—choice probabilities, yes/no probability, or ordinal expected values—without autoregressive generation, achieving 94.71% test and 86.02% OOD hard accuracy.
Open-Jev-27B-v1.1A 15.5M-parameter LoRA adapter and scalar decision head for Qwen3.8-27B that scores supplied candidates and returns typed choice, yes/no, and ordinal-score probabilities without autoregressive generation, scoring 85.28% on 231 public JevBench tasks (72.07% Hard).
clef-text-0.6b-coremlCore ML packages (fp16 decoder + joint schema head, ~1.3 GB) running a 0.70B text decision model distilled from Cloudflare's 9B clef-flash into Qwen3-0.6B, executing fully on Apple Neural Engine or GPU to return typed probabilities for `noul`/`choice`/`score` questions in one ~35–77 ms forward pass at 79.1% average gold accuracy.
Winnow-E4BWinnow-E4B is a compact merged-GGUF Gemma 4 E4B fine-tune that scores supplied answer options for yes/no, choice, and score questions over a shared state—returning probabilities without generating text—while keeping chat and image support for fast local decisions.
Jeff-1Jeff-1 is an Apache-2.0 LoRA adapter for Qwen3-4B-Instruct-2507 that scores caller-supplied labels to return typed decisions—choice, yes/no, and ordinal scores with probabilities and confidence—for local inference tasks like fact-checking without parsing generated text.
SevSev is a Gemma-4-E4B-it LoRA decision model that classifies text against arbitrary plain-language policy questions in one non-generative forward pass, returning answer probabilities from embedding similarity for safety, injection, legal, and tax guardrails.
Sieve-31BSieve-31B is an Apache-2.0 rank-128 LoRA adapter on Gemma-4-31B-it that reads a text or JSON state plus typed choice/noul/score questions and returns a calibrated probability for every option without generating text, reaching 62.52 on the public Decision Index 0.3.
Jev-Omni-Q4_K_M-GGUFA Q4_K_M GGUF quantization of the Jev-Omni 12B multimodal decision classifier that, together with its original-precision projector and FP32 decision head, lets llama.cpp return option probabilities over text, image, audio, and video inputs on CPU or GPU.
jeevesJeeves-9B is a fused Qwen3.5-9B decision model that writes a reasoning chain for each yes/no, multiple-choice, or rating question about a given state and returns calibrated per-option probabilities, bundled with a pointer head, fitted temperature, and two diffusion drafters for faster serving.
On-Fly-JevOn-Fly-Jev is a 96M-parameter PyTorch typed-decision model built on a spiking neural network trunk of 100 real Drosophila MaleCNS connectome neurons, returning calibrated probabilities for choice/score/noul questions in a single forward pass at ~137 decisions per second (7.3 ms p50) with best-in-group 0.045 ECE.
basal-1.0-4.5Bbasal-1.0-4.5B is a 4.5B-parameter Polish/English model that answers typed questions (choice, yes/no, score) about a state by returning calibrated per-option probabilities in one forward pass, enabling fast dynamic classification with confidence-thresholded auto-acceptance.
jev-schema-scorer-deberta-v3-largeA fine-tuned DeBERTa-v3-large encoder with a scalar head that scores (state, question + candidate) pairs, decoding grouped logits via softmax into choice, boolean (noul), and ordinal score answers—reading criteria and option ids at inference time so one checkpoint handles new label sets without retraining.
Sieve-31B-v3Sieve-31B-v3, a full fine-tune of Gemma-4-31B-it, reads a text or JSON state once and returns calibrated probabilities for every option of typed choice/noul/score questions via bidirectional letter-logit readout—never generating text—scoring 71.68 on Decision Index 0.3.
ArseneLupin-v1.1ArseneLupin-v1.1 is an Apache-2.0 decision model fine-tuned (merged LoRA) from Qwen3.5-4B that returns calibrated probability distributions for typed yes/no, choice, and scored questions over a given state in a single forward pass per option, served via a System One-style API in full bfloat16 (transformers) or 8-bit GGUF (llama.cpp) form.
Blink-v0.3-26B-A4B-NVFP4Blink-v0.3-26B-A4B-NVFP4 is a Gemma-4-based MoE decision model (4B active params, NVFP4, 32k context) that returns calibrated per-option probabilities in a single forward pass, replacing "ask an LLM and parse prose" calls for routing, moderation, tagging, gating, dedup, and yes/no checks.
jebadiah-9b-v2Jebadiah-9b-v2 is a merged bf16 LoRA finetune of Qwen3.5-9B that answers typed choice, yes/no, and rubric-score questions by returning calibrated probability distributions over option labels instead of generating text, served via AINode's /v1/decide and /v1/systemone routes.
rsi-jev-v6.1-vl-27bA 26.66B-parameter typed-decision vision-language model—a weight average of two Qwen3.8-27B fine-tunes—that returns calibrated per-option probabilities for choice/noul/score questions over text and up to four images in one forward pass, early-exiting at layers 48/56/64.
basal-1.0-1.5Bbasal-1.0-1.5B is a 1.5B-parameter Polish/English decision model distilled from basal-1.0-4.5B that, given a state and a typed question (choice, yes/no, or score) with options defined in plain language, returns calibrated per-option probabilities in a single forward pass — enabling dynamic, retraining-free classification, routing, and triage at ~2× lower latency than its teacher.
jev-berta-base-zeroshot-classifierDeBERTa-v3-base model that scores variable-length candidate choices against a context and query, using candidate-set self-attention so each option is judged conditioned on all alternatives, and a JEV adapter mapping state, instructions, and criteria to predictions per question.
hotdog-27Bhotdog-27B is a rank-16 LoRA adapter on Qwen3.8-27B that answers yes/no questions about text with a probability and binary decision read directly from logits, achieving 98.28% independent-test accuracy and strong calibration without generating reasoning.
jev-9b-coremlA Core ML build of JEV-9B for Apple silicon that makes typed decisions over text and images—yes/no, 0–5 scores, or 2–256-option choices—returning a calibrated probability for every option in a single GPU forward pass, callable via Swift's `JevManager`.
Qwen3.5-4B-Judge-1MA 4B-parameter judge built on Qwen3.5-4B with an Unsloth LoRA decision head that outputs calibrated probabilities for fixed-choice decisions such as RAG claim support, safety checks, and task/domain routing, trained on 39,592 rows across 20 task families.
vjev-visionvjev-vision is a Qwen3.5-4B-based listwise decision model that, given text and image states plus typed questions (noul, choice, score), returns calibrated probabilities for every option in a single forward pass without text generation.
gliner2.5-cua-grounder-macos-v1A merged FP32 GLiNER2.5 fine-tune that chooses an action ID, `reobserve`, or `abstain` from a bounded action menu for local macOS CUA automation—without screenshots or tool calls—and is verified on browser DOM and native AppKit fixtures with low-latency MPS inference.
jev-control-corejev-control-core is a fully fine-tuned 0.8B decoder that answers typed choice/score questions at embedded agent-harness decision sites (tool routing, injection gates, triage, escalation) in one restricted-option forward pass, yielding calibrated confidence at ~18.5ms per decision.
laya-groundedA fine-tune of the laya typed-decision model that improves grounding on explicitly stated facts, consistency between logical inverses, and probability calibration, at the cost of worse phishing detection and option-name routing stability.
openjevxOpenJevX is a small CPU-first ONNX decision model (8-bit, 598 MB) that answers yes/no, pick-one, and 0–N rating questions about a JSON or text state with per-answer probabilities and confidence, running fully locally via Docker or Python at ~22–540 ms CPU latency.
PurpleMIST-Flash-1.0PurpleMIST-Flash-1.0 is a 7.9B-parameter decision model that answers typed choice, yes/no, and score questions about unstructured state with calibrated probability distributions in a single forward pass, in English and eight African languages.
rsi-jev-v6.0-vl-4brsi-jev-v6.0-vl-4b is a 4.69B-parameter vision-language decision model fine-tuned from Qwen3.5-4B-Base that answers typed questions (yes/no, multiple-choice, rubric scores) about text plus up to four images in a single reasoning-free forward pass, returning calibrated per-option probabilities via early-exit heads at layers 16, 20, and 32 for adaptive compute.
tasksource-jev-nano-v0Tasksource-JEV-Nano-v0 is a 149M-parameter ColBERT-style late-interaction model (fine-tuned from LateOn on 512k typed decisions) that picks the best option from a list—choice, yes/no, or graded scoring—with a cacheable situation encoding and permutation-equivariant option scoring instead of a classifier head.
ultrabert-jev-380m-softmax-yesnoAn UltraBERT-380M encoder fine-tuned to pick the correct option in a single forward pass—scoring each mask-prefixed option by its "yes"−"no" logit—achieving ~73% accuracy across 13 datasets, provided fp32 weights and the exact training prompt are used.
ArseneLupin-mini-v1.0ArseneLupin-mini-v1.0 is a 0.87B-parameter decision model (Qwen3.5-0.8B + merged LoRA and a trained read-out head) that returns calibrated probability distributions for yes/no, choice, and score questions over any input state in one forward pass per option, with no text generation, served via a System One-compatible API in bf16 or GGUF Q8_0.
AutoTrust-JEV-27B-v1.1-mlx-8bitAn 8-bit MLX build (28 GB) of AutoTrust-JEV-27B-v1.1 for Apple Silicon, a decision model that answers each call—a choice among up to 255 options or yes/no—with per-option probabilities in one bidirectional forward pass, matching the bf16 original on all 231 test decisions.
jebadiah-9b-v0Jebadiah-9B-v0 is an Apache-2.0 LoRA adapter on Qwen3.5-9B-Base that answers typed decision questions (choice, yes/no "noul", rubric score) by emitting a calibrated probability distribution over option-label tokens instead of generating text, trained on 11,072 public questions and served via AINode's /v1/decide and /v1/systemone routes.
JEVisionJEVision is a Qwen3.5-0.8B-based System One decision stack with text and visual LoRA adapters, pointer heads, and a local runtime that accepts text plus optional images and returns typed Choice, Noul, or Score answers through one API.
mario-student-coremlA 9.0M-parameter MobileNetV3-Large pixel policy (fp16, 17 MB Core ML package, ~0.5 ms/decision on Apple Neural Engine) that plays Super Mario Bros. on NES, trained via imitation learning and DAgger from a beam-search solver to choose among 11 eight-frame moves, clearing 78/96 level-offset combinations with emulator lookahead.
MiniCPM-OCR-Decision-2B-LiteRTLiteRT conversion of MiniCPM-OCR-Jev, a 2B vision model that scores OCR candidates 0–1 by comparing a region image, its text, and a rendered image, with bucketed TFLite graphs for desktop and Android GPU/NPU plus a PyTorch-free Python host.
Qwen3.5-4B-JudgeA 4B-parameter judge (Qwen3.5-4B + Unsloth LoRA decision head) that returns calibrated probabilities for RAG groundedness and safety/refusal decisions, scoring 87.9% and 90.7% on in-domain tests and 78.4% on unseen MedHallu, all above Jev 1.14's zero-shot results.
Rook-V1Rook-V1 is an uncalibrated self-hosted research decision model for bounded choices, routing, and candidate scoring, providing a LoRA adapter, decision head, tokenizer, configs, reload code, integrity manifest, and evaluation artifacts for reconstruction on the pinned Lux 9B.
rsi-jev-v2.0-qwen3.5-2brsi-jev-v2.0-qwen3.5-2b is a Qwen3.5-2B typed-decision model that answers yes/no, pick-one, and rubric-score questions about documents in one forward pass, returning calibrated per-option probabilities for trustworthy automated decisions and escalation routing.
safe-layasafe-laya is an Apache-2.0 421M ModernBERT decision model that scores untrusted prompts, retrieved docs, and tool outputs for prompt-injection/jailbreak risk in one ~30ms local pass—a cheap, private injection-guard signal with signed AI-BOM and reproducible training.
serin-4bSerin-4B is a 4-bit-runnable QLoRA adapter on Qwen3.5-4B that acts as a local LLM-as-judge for Polish and English chatbot security audits, returning per-option probabilities for typed questions like "did the attack succeed?" from a single forward pass without generating text.
Solar-Decision-JEV-27B-V2-GGUFA GGUF quantization ladder (IQ4_XS through Q8_0, ~10–27 GiB) of the Solar-Decision-JEV-27B-V2 merge for fast llama.cpp chat with speculative-decoding drafter tensors preserved, though calibrated typed decisions still require the bf16 weights in vLLM.
ultrabert-jev-380m-sigmoid-normAn UltraBERT-380M encoder fine-tuned into a single-forward-pass decision model that picks one option from a list by comparing "yes"−"no" logits at mask tokens placed before each option (≈72% accuracy on 13 reasoning datasets), plus a token-level evidence-scoring head.
ultrabert-jev-380m-softmax-yesA 380M-parameter UltraBERT encoder that answers multiple-choice questions over up-to-10,240-token inputs in a single forward pass by softmax-scoring each option's mask " yes" logit (68.7% mean test accuracy on 13 reasoning datasets), with an optional head scoring evidence tokens.
ultrabert-jev-380m-sqrt-softplusAn UltraBERT-380M encoder checkpoint that selects one of several options in a single non-generative forward pass from mask-position "yes"−"no" logits, trained with square-root-softplus normalization and reaching 73.2% dev / 71.9% test accuracy across 13 reasoning datasets.
verdictVerdict is a small trained head over frozen SigLIP 2 that answers typed image questions—multiple-choice with an "other" abstention and true/false statements—in a single forward pass, returning calibrated per-option probabilities with no text generation.
ArseneLupinstral-v1.0ArseneLupinstral-v1.0 is an Apache-2.0, 8.92B-parameter Mistral Ministral-3-8B decision model whose merged LoRA and trained readout head turn a supplied state and yes/no, choice, or score questions into calibrated probability distributions—one forward pass per option, no generated text—via a System One-style API.
ballotballot-E-v0 is a ~149M-parameter ModernBERT fine-tune that returns per-option probabilities for multiple-choice, yes/no, or rating questions in one non-generative forward pass, structurally guaranteeing order-invariant answers (0.0% flip rate) at 68% held-out accuracy.
ballot-jev-0.5bballot-jev-0.5b is a 506M-parameter CPU-first typed-decision model that returns calibrated full probability distributions for choice, true/false, and ordinal questions, and is structurally permutation-invariant—reordering options provably cannot change its answer.
Blink-v0.5-26B-A4B-NVFP4An NVFP4-quantized 26B-A4B MoE model that turns a state, question, and options into calibrated per-option probabilities in one forward pass, replacing LLM calls for routing, moderation, tagging, and gating, with optional think-when-unsure reasoning.
brainsurged-open-jev-qwen3.5-0.8bA 516.8 MiB Q4_K_M GGUF of Qwen3.5-0.8B, tuned with Rank-16 LoRA and zero-gradient roof calibration for low-latency, schema-constrained triage and routing decisions, with optional Jump-Forward CUDA serving sources and evaluation records.
brainsurged-open-jev-qwen3.5-4bA Q4_K_M GGUF build of Qwen3.5-4B tuned via rank-16 LoRA and zero-gradient roof calibration for schema-constrained JSON decision-making, with archived patches, llama.cpp/Ollama deployment paths, and a CUDA jump-forward serving stack benchmarked on Jevlab tasks.
chessjev-minilm-v1A chess option scorer that finetunes MiniLM-L6 with an attention head to emit logits for all legal moves from a FEN context in one forward pass, distilled from Stockfish 19 (0.445 teacher agreement, 0.667 on tactics) for fast move ranking rather than verified playing strength.
clef-27b-mps-nf4Clef is a 27B multimodal model post-trained from Qwen3.8-27B that turns a state (text, JSON, images, or video) and a schema of typed questions into per-option probabilities for all questions in one forward pass, with no free-form generation or parsing.
embeddinggemma2-decision-routerAn open-source routing model that maps text, image, or audio workflow states to typed, calibrated decisions (choice/score/boolean) in one forward pass via a frozen EmbeddingGemma 2 backbone plus a 3.55M set-equivariant head, offering zero option-order bias, out-of-scope abstention gating, and ~54 ms latency.
geppoGeppo is an Italian semantic-decision model pairing a frozen Emma-6 causal backbone with a 262K-parameter pointer head that scores candidate descriptions to answer choice, yes/no, and ordinal questions locally via a standalone runtime with calibrated probabilities.
gev-e2bGev-e2b pairs a LoRA adapter and pointer head on Gemma 4 E2B to score options for choice, yes/no, and ordinal questions, outputting one answer with a calibrated probability distribution (82.9% trained-source, 62.5% new-source accuracy).
jev-27b-coreaiA 4-bit quantized Core AI conversion of JEV-27B that runs locally on Apple silicon Macs, returning a probability for each allowed answer to yes/no, rating, or 2–16-option choice questions through a Swift CLI, with a checksum-verified bundle and documented fidelity measurements.
jev-control-esA 149M-parameter ModernBERT-base decision model that scores each option at its own dedicated [MASK] token in a single bidirectional pass, returning calibrated softmax probabilities for agent-harness decision sites (injection gates, tool routing, context ranking, sufficiency, escalation) at ~7.7 ms per decision.
jev-spam-decision-2bA LoRA adapter for Qwen3.5-2B that answers "Is this forum message spam?" by extracting a calibrated P(yes) from next-token logits over `yes`/`no`, reaching 0.9993 accuracy/F1 and 0.001 ECE on the Deysi spam-detection test split, with tuning and usage details included.
JevAny-Robotics-4B-LoRAA 4B LoRA on Qwen3.5-4B that judges robot manipulation in one forward pass, returning calibrated per-option probabilities for task progress, rollout scores, success, failure mode, and trajectory preference over real/simulated robot imagery, near-top across evaluated benchmarks.
jevons-lfm25-1.2b-systemoneA LoRA adapter for LiquidAI's LFM2.5-1.2B-Instruct (8-bit MLX) that biases outputs toward option-key, yes/no, and score-digit tokens so the jevons server can logit-score and assemble System One decisions, with a 72.4% holdout modal accuracy at T=1.
JJ-1-2BJJ-1-2B is a compact Japanese/English decision-classification model that returns choices, yes/no probabilities and scores from text or JSON—reaching 94.63% on business workflows—with BF16/Q8_0 weights plus Transformers, CLI/HTTP and llama.cpp inference code.
llm-jev-qwen3-0.6bllm-jev-qwen3-0.6b is a Qwen3-0.6B model fine-tuned with a LoRA adapter (merged weights included) that scores numbered answer options via a candidate trie under the LLM-as-Jev method, raising held-out option-selection accuracy from 46.7% to 65.0% over the base model.
llm-jev-qwen3-4bA LoRA fine-tuned Qwen3-4B decision model (merged weights plus adapter) that scores numbered answer options for a given document and question through a candidate trie rather than generating text, raising held-out option-selection accuracy from 72.2% to 79.3% while leaving external benchmarks essentially unchanged.
mati-14bMati-14B is a Kev-compatible LoRA adapter and calibrated pointer head for Qwen3-14B-Base that returns temperature-calibrated probability distributions over declared options for decision and ranking workloads.
nomaAn open decision model on a truncated Qwen3.5-4B backbone that returns calibrated per-option probabilities plus abstain and uncertainty for typed questions in ~16 ms, enabling fast routing, triage, scoring, and agent-step verification without text generation.
OpenVerdict-4BOpenVerdict-4B is an Apache-2.0 LoRA fine-tune of Qwen3.5-4B that returns calibrated probabilities for every option of a stated decision question—via one forward pass or a ~50-token chain-of-thought—across text and image inputs for routing, policy checks, moderation, and tool-use decisions.
Qwen3.8-Flash-Next-Hybrid-DecisionQwen3.8-Flash-Next NVFP4/FP8-PLE weights packaged with a patched vLLM runtime that adds a typed decision API returning masked-softmax probabilities over 2–256 caller-supplied options or booleans, while serving ordinary chat on the same engine instance.
rsi-jev-v4.0-vl-qwen3.5-2bA 2B typed-decision model built on Qwen3.5-2B-Base with a trained option-scoring head that returns calibrated per-option probabilities for yes/no, pick-one-of-k, and rubric-score questions about a document plus up to four images in a single forward pass, generating no text.
Sev-4BSev-4B is a Qwen3.5-4B LoRA adapter with a decision head that scores candidate answers about security logs and recovered programs—including authorization-policy checks grounded in real code—giving analysts calibrated evidence review (84.0% held-out policy accuracy) without generating text.
spark-s1-1.7b-v3spark-s1-1.7b-v3 is a LoRA-tuned Qwen3-1.7B decision model that answers typed questions (choice, boolean, score) over request-time options by softmaxing first-token logits into a probability distribution and confidence in one ~30 ms forward pass, with no text generation.
spark-s1-4b-v3spark-s1-4b-v3 is a Qwen3-4B LoRA fine-tune that answers typed questions (choice, boolean, ordered score) about a rendered state by softmaxing first-token logits over option letters—returning a calibrated per-option probability distribution plus confidence in one forward pass with no text generation—for fast local agent decisions.
spark-s1-4b-v6spark-s1-4b-v6 is a LoRA-merged Qwen3.5-4B System-1 decision model that, given a state and a typed question, returns calibrated probabilities over lettered options in one forward pass — useful for fast local agent guardrails such as allow/ask/deny tool-call decisions.
StandardOne-3B-SHStandardOne-3B-SH is a Ministral-3 3B model with a joint schema head that scores every option of every question (choice, yes/no, ordinal) in one pass over up to 32k tokens, returning calibrated probabilities without generating text.
vera-deberta-v3-largeVERA is a fine-tuned DeBERTa-v3-large model that scores and ranks candidate headline or short-copy variants by predicted click rate in a single forward pass, trained on 62,695 arms from 32,487 real A/B tests to pick higher-performing variants without generation or prompting.
WebJev-35B-A3B-GGUFGGUF quantizations (Q4_K_M, Q6_K, Q8_0) of WebJev-35B-A3B, a decision model that scores typed-choice options for browser agents in one forward pass, plus a llama.cpp adapter exposing the /v1/systemone route so it runs on home hardware instead of vLLM on an 80 GB GPU.
zeiger-0.6bZeiger-0.6B is a 614M-parameter encoder-decision model (Qwen3-0.6B backbone plus 18M scoring head) that, given a web page's element descriptions and an instruction, picks the target element—or answers "none"—in one calibrated 32k-context forward pass at up to 1,501 candidates, enabling fast, reliable browser-agent element grounding without generation or chunked tournaments.
zeiger2-3.5-0.8bZeiger2 3.5-0.8B is an experimental English/German decision model that encodes a web page, instruction, and answer options in one pass and returns calibrated per-option probabilities — for element grounding, yes/no, ordinal, and "none of these" questions — scoring 0.838 on held-out web tasks and 17.6 on the Decision Index.
OneJevOneJev is a multimodal System One decision model (0.8B–27B) that reads screenshots, photos, video, or text and returns calibrated probabilities for every option of every question in one shared forward pass, serving TypeSafe's System One API for fast GUI-agent decisions.