Everything we found on Jev-style decision models: the papers, the open models, the leaderboards, and a code-level audit of 27 implementations. Jev appeared on 15 September 2026. By mid-October about 130 papers mentioned it, and nearly 2,000 Hugging Face models were tagged as decision models.
Numbers are as reported by each source; claims taken from a README, model card or notes rather than a paper table are marked (self-reported).
One forward pass, one calibrated distribution.
Jev is a closed "System One" decision model from TypeSafe AI, launched 15 September 2026. It answers a typed question about a state with a calibrated probability, without generating text. The open work below copies its interface.
Pick one of the declared options. "None of these" is not a primitive; some open models add it as an option or a separate gate.
Place the input on a declared scale, for example 1–5 relevance.
Probability that the answer to a yes/no question is yes.
schematic: the glyphs show the output type, not data
Five papers to read first.
The five we found most notable, then the rest. All appeared after Jev's 15 Sep launch, so citation counts are not yet meaningful.
Request and each option are encoded separately (options cacheable). A small decision block lets the request vector read the whole option set before scoring, and a coverage gate gives an explicit "none". Supervised plus filtered distillation, no RL. Run at 2,048 options.
First open version of Jev's RL, on Qwen3-1.7B. Rewards the committed answer distribution with a strictly proper scoring rule: calibrate first, then reinforce. Beats SFT, RFT and GRPO on selective prediction; cannot beat CE when uncertainty is annotator disagreement.
Read option-letter logits from open LLMs, then LoRA fine-tune. Matches or beats Jev on decision tasks. Its ablation: LoRA drives the gain, not the head (Gemma 4 31B, Capability on OpenDecisions: head only 78.9, LoRA 84.9).
The main independent evaluation of Jev: 37 datasets, calibration, comparisons with open Qwen and Gemma models.
Numbered options read through a prefix-free trie (unbounded K); KL anchors stop generation from degrading. Paper only.
Training-free readout from any LLM; averaging over option rotations cuts answer flips from 0.230 to 0.073. Code → evidence
Label-free decision head on a frozen LLM, distilled from the full model; 75.0 vs 75.7 using 24 of 36 layers. Code
Self-distilled decision models; finds they pick the likely outcome but overstate its probability.
Order invariance for multiple-choice decisions.
Jev used as policy, score rater, meta-regulator and transition model inside RL loops.
Laya-family encoder for Chinese with GRPO-style RLCD and DDP training. Code
RLCD-trained decisions used as a zero-shot detector.
A paper benchmark for typed decisions: calibration, sensitivity to wording, cost.
Beyond text: one video grounder, one bi-encoder, image-side speed-ups.
Everything Jev-style we found beyond text.
4B and 35B-A3B models. One sequence [video → question → one timestamp option per 2 frames → decide token]; the decide token scores every option. Independent "inside the event" probabilities, BCE plus soft-IoU loss, Kadane decoding. Average mIoU 65.3 (4B) on Charades-STA, ActivityNet and QVHighlights. Short and medium clips only.
A 30B decoder used as a bi-encoder: request and candidates encoded separately, text plus image and video, for decisions and million-scale retrieval. Evaluation code only.
Small VLMs (2B, 4B) making Jev-style choices on images; object and action selection left as future work. Paper only.
The closed Jev API picks phone-UI actions; a VLM plans only occasionally. −73% cost.
A fast decision model hands over to a reasoning VLM when a gate opens (Doom).
Remote-sensing classification with a calibrated "abstain" option.
Agriculture candidate scoring with adapted Qwen3.5 models.
Zero-shot video quality prediction with a decision model.
A visual verifier whose yes-probabilities are GRPO rewards for image generation.
Robustness benchmark for decision models, text and vision.
Text-only weights go down to 17M; multimodal starts at 0.6B.
Weights you can download, multimodal first. Video support means whole-clip question answering at 2 fps.
Text, images, video at 2 fps (max 32 frames). #1 on the Jev Decision Index text board (66.3) and its Vision Index (71.3)(self-reported). Noncausal attention; trained on AMD MI325X. audit ↗
0.6B: text, image, speech, no video. d1-3B: text and images.
Letter readout with option-count temperature calibration. audit ↗
4B, text and image; #1 open-weight on JevBench 7–15 Oct (72.5), per its model card(self-reported). The live board has since moved (Quyet-1.0-Large leads; see leaderboards).
27B / 9B / 35B. clef-omni reads text, image, audio and video; a joint schema head scores all options of all questions. No training code. audit ↗
27B, image; distilled from Jev outputs.
Sentence-Transformers based, runs in the browser; second on S1MB behind Jev in the board's 30 Sep snapshot(self-reported). audit ↗
Laya architecture on mmBERT-small; weights and inference only.
ModernBERT / mmBERT encoders; the most-starred open System-One repo. audit ↗
ModernBERT-large, CPU-served under 15 ms(self-reported); options masked from each other. audit ↗
Qwen3-0.6B trained on game decisions (ViZDoom, Maze, Snake). audit ↗
Order-invariant candidate set encoder on Qwen3-0.6B. audit ↗
Pointer head from a decide token; best-engineered open repo. audit ↗
Letter SFT through a Together API job; uncalibrated by its own README(self-reported). audit ↗
Letter readout, per-option-count temperature, GGUF and vLLM serving. audit ↗
Cross-encoder yes/no per option, CE plus Brier, one fitted temperature. audit ↗
Letter readout with a calibrated knockout for 100+ options; #2 of 76 on JevBench v1.4(self-reported). audit ↗
Two-tower contrastive heads; cleanest cacheable reference. audit ↗
Eleven demos worth trying.
We found 282 live demos of Jev-style models (all checked on 16 Oct 2026). These eleven are the most carefully made and show the most: probabilities, calibration, abstention, latency, the one-pass mechanics, and two robots in simulation (one live, one recorded). Screenshots 16 Oct; click one to open the demo.
Support triage, guardrails, routing and more, each answered in one pass with a probability per option; the thresholds live in your code, not in the model.
Jev's /v1/systemone request format with both sizes side by side, plus probes for question isolation and option-boundary forgery.
Encode the state once, decide in parallel: cached vs naive scoring vs JSON generation, with forward-pass counts.
Switch between open decision models on the same input, on your own device, with latency and forward passes counted.
Calibrated distributions in one pass; an Escalate tab hands low-confidence questions to System 2 thinking with the same weights.
Photo-against-record checks with an explicit can't-tell probability, so you can trade coverage for accuracy.
A probability for every answer on text, images and video; models side by side with latency; boxes scored on a zooming grid.
Eleven camera and drawing mini-games, each driven by a few typed questions answered in one pass.
Temporal grounding and video QA in one pass, with a probability per timestamp; several questions share one video encoding.
A simulated Pollen Microduck sees only its own camera; each step the model picks forward, left, right or kick, about 70 ms per decision, on the action space of a published eval.
Each step the model reads the top camera and answers two questions, is the target left or right of the gripper and above or below it; the arm grasps the cube and drops it in the tray. 75% of 20 random scenes, about 240 ms per decision, per the model card (self-reported).
to try closed Jev itself: TypeSafe playground (account) · OpenRouter playground (credits)
Five boards, five definitions of good.
Where decision models are ranked. Screenshots taken on 16 October 2026, except jev-bench, which shows the board's own chart; click an image for the live board.
| Board | Version · size | Measures | Leader, as stated | Jev, as stated |
|---|---|---|---|---|
| Decision Index | v0.3.2 · 168 models | 37 public benchmarks + private tests · text and vision boards | vLLM-SR d3 27B⁴ 66.3 full score | Jev 1.14 61.5 full score |
| DecideBench | v1.1 · 400 decisions | accuracy · cost per M tasks · latency | overall leader: a chat LLM¹ | JEV 98.0% accleads the decision models |
| System One Mosaic Benchmark | 165 models · 137 benchmarks | text only · Borda score over Noul, Choice, Score | Superion S1 91.21 Borda | Jev 1.14 Preview² 81.85 Borda |
| JevBench | v1.7.46 · 162 ranked | intelligence · calibration · cost · latency | Quyet-1.0-Large 81.7 capability | Jev 1.13.0 77.1 capability |
| jev-bench | v0.1.1 · 22,773 records | accuracy vs calibration error, human labels | open readout-fine-tuned models³ | Jev 1.13.0 0.733 acc · 0.113 ECE |
The main board for open Jev reproductions: 37 public benchmarks plus private tests, with separate text and vision boards. Top on 16 Oct: vLLM-SR d3 27B (66.3); Jev 1.14 at 61.5.
Accuracy, cost per million tasks and latency, decision models next to chat LLMs. JEV leads the decision models at 98.0%. Code
Text only, ranked by Borda score across Noul, Choice and Score tasks. Superion S1 leads (91.21). Write-up
Intelligence, calibration, cost and latency, run independently of TypeSafe. The capability score averages intelligence and calibration over systems within 2× Jev's cost and latency. Quyet-1.0-Large leads (81.7); Jev 1.13.0 scores 77.1 as the reference.
Every model on the same human-labelled test set, accuracy against calibration error. Jev 1.13.0: 0.733 accuracy, ECE 0.113; readout-fine-tuned open models beat it on both. Chart from the dataset card (2 Oct).
Unranked collections of models and papers.
Places that collect decision models and papers without ranking them.
Hugging Face tag Every model its authors tagged as a decision model.
Hugging Face search Repos with "jev" in the name, most liked first.
systemonemodels.ai A directory of decision models with self-reported scores(self-reported), which are not comparable across models.
arXiv search Papers mentioning Jev, 19 Sep – 14 Oct 2026.
22 of 27 audited repos encode request and options jointly.
We read the code of 27 implementations, all scored the same way. The first question for any design: are the request and the options encoded together in one sequence (joint), or separately (cacheable options)? Stars as of 15 Oct 2026.
| Repo | Encoding | Decision head | None | Max options | Stars · 15 Oct | Take-away |
|---|---|---|---|---|---|---|
| Text, decision heads on LLM backbones· 5 | ||||||
| Jeva | separate | Request vector reads the option set (attention), then dot product | gate | 2,048 tested | 76 | Only design built for cacheable options; head is 470 lines of plain PyTorch |
| Kev | joint | Decide-token pointer over option tokens | no | 255 | 9,178 | Best engineering (447 tests, ROCm); notes record that isolation and head width did not help(self-reported) |
| NanoJev | joint | Per-option scalar + set attention | no | 255 | 2,552 | Proper-score RL estimator, found no better than CE(self-reported) |
| Decider | joint | Letter-token logits | no | 255 | 1,147 | Best calibration practice: T(n) = a + b·ln n |
| AgentJev | joint | 2-layer set transformer over candidates | no | 255 | 386 | Clean candidate set encoder (~60 lines) |
| Text, encoder models· 7 | ||||||
| Laya | joint | 2-layer head, MLP at each option marker | untrained | 32 | 33,022 | Huge community; "RLCD" loss and per-K temperatures; shipped ECE 0.57(self-reported) |
| GLiNER2.5-Decide | joint | MLP per label token | threshold | 512 positions | 2,374 | Excellent engineering; options are tokens in a 512-position encoder |
| Von | joint | MLP at option markers; options masked from each other | no | 8K tokens | 906 | Order-invariant by construction; CE + 0.5·Brier |
| Bekko | shared prefix | Linear + optional set attention | no | no cap | 37 | Cleanest small code; set head marked experimental |
| Chinese-Jev | joint | Laya head | act head | 32 | 16 | Compact RLCD + DDP |
| Verdict | separate | Cosine or logistic regression | conformal | no cap | 3 | Right towers, no real head |
| Bongard | joint | Bilinear candidate × decision token | no | — | 3 | 7.5B, causal, no tests |
| Text, recipes and two-tower models· 9 | ||||||
| CLM | separate | Cosine of two MLP heads | no | no cap | 3,033 | Clean two-tower; loses to joint heads (44.3% vs ~75%, FunDM's audit) |
| SemIf | joint | Letter logits, no training | no | 16 | 4,829 | Interface demo |
| AnyJev | joint | Letter logits, rotation averaging | no | 26 | 1,276 | Quantifies order bias → evidence |
| Open-Jev | joint | Yes/no scalar per option | no | K passes | 415 | CE + Brier, one temperature |
| StartLux | joint | Letter logits, bidirectional prompt | option | 26 + knockout | 415 | Explicit "none" costs no accuracy → evidence |
| Tev1 | joint | Letter SFT | no | 24 | 236 | A data recipe |
| JevK5 | joint | Letter logits + knockout | no | 16 per pass | 155 | Only calibrated recipe measured at 151 options |
| FunDM | joint | One-layer set transformer over option states | no | 10 tested | 0 | Most liftable set head (~25 lines) → evidence |
| SII-Decision | joint | First-token logits | no | 255 | 3 | Self-distillation recipe |
| Vision and video· 6 | ||||||
| VideoJev | joint | Decide-token pointer over timestamp options | no | 1 per 2 frames | 123 | Best video evidence; soft-IoU +4.6 mIoU; CUDA-only kernels → evidence |
| OmniJev | shared prefix | FiLM-gated scorer per option | abstain | — | 163 | Proper-score losses; options fully isolated |
| OmniMed-Jev | joint | Linear + bilinear + set attention | no | any | 1 | Query-conditioned set head (~70 lines) |
| Clef | joint | Options route evidence from a memory, questions attend to each other | no | — | HF only | Richest head; no training code, no ablation |
| Visual Jev | joint | LM-head letters | experimental | 16 | 46 | Typed head tied the LM head |
| vllm-sr d3 | joint | 255-code readout, noncausal | no | 255 | 6,180 | Only one verified on ROCm; no head training code |
Eight measured comparisons, each on its own metric.
No repository ablates its own head design. These are the measurements that exist, each from the cited source.
each plot: own metric, own axis from zero · ○ before ● after · dashed = README(self-reported)
Putting the options into the request lifts MMLU from 56.5 to 74.6. FunDM finds +16–17 MMLU-Pro for joint over option-only encoding.
A set-attention layer on top of joint states.FunDM, Table 4 · +0.4 / +2.1 mean accuracyStartLux · up to +2.4 Decision Index (bidirectional attention in the prompt)
FunDM, Table 4Gemma 4 31B. Training only the head: 78.9. LoRA with the head frozen: 84.9. LoRA ranks 16–128 score about the same.
Adding a duration-weighted soft-IoU loss to per-timestep BCE: 60.7 → 65.3 mIoU, short clips.
Pre-train then hard negatives 69.2, vs 52.1 without and 62.4 with hard negatives from the start. Replay keeps 68.5 vs 56.2.
Adding a "None of the above" option left accuracy unchanged (84.66 both ways, 51,506 questions).
StartLux-DecisionAveraging over option rotations cuts answer flips under reordering from 0.230 to 0.073; ECE 0.240 → 0.184.
Kev's notes list option isolation, head width and permutation-KL as tried without help. Visual Jev's typed head tied the LM head (0.761).
Kev PLAN.md(self-reported)Jev's RLCD is unpublished; here is what is open.
What exists for training calibrated decisions with RL. Jev's own RLCD is unpublished.
Proper-score reward on the committed answer distribution, leave-one-out baselines, KL anchor; estimators tested for unbiasedness. Text only.
Rewards the improvement in probability quality after reasoning, minus a length penalty. RL training code not released.
Noise-perturbed logits rewarded by proper scores plus soft CE: a regulariser around a fixed target, not RL against an environment.
Paired proper-score policy gradient with exact gradient checks; its own study finds it no better than direct CE or Brier(self-reported).
PPO with a log-score belief reward, documented in docs/RL.md; the training code is not in the repo.
RLCR (Brier reward), Rewarding Doubt (log score). Not decision models.