The open decision-model spectrum runs from a 0.5B prototype to a 35B mixture-of-experts model, and the ranking does not follow parameter count. A small pointer-head model can be right when the state is short and the labels familiar. A 27B can beat a 35B on calibration while losing on raw accuracy. Compare within a training family and on a fixed task, because published numbers use different suites, judges, and tuning.
What size actually buys
Size buys capacity: more layers, more parameters, more tolerance for unfamiliar question types. It does not buy calibration or a clean decision boundary; those come from data and the head. Kev-0.6B is the clean experiment. It scores 0.801 on decision-v4 development and 0.808 in-distribution on a locked test, but 0.620 on transfer-v4 development and 0.642 out of domain. The card says transfer is flat at about 0.60 across eight hyperparameter mutations and three seeds, so capacity, not tuning, is the bottleneck.
The larger sibling tells the other half. Kev-4B reports 0.838 transfer-v4 accuracy with Brier 0.224, 0.803 on hard-v1, and 0.756 on devtools-v1, all card-reported. That is a 14-point gain over Kev-0.8B's 0.697, with Kev-0.6B lower still. The jump is not linear in parameters and not a guarantee on your task; it is evidence that the 0.6B class is capacity-limited out of domain.
The JEVal benchmark adds a broader caution. Across 11,257 bilingual instances from 36 datasets and 10 domains, decision models are competitive when evidence resolves the decision, but overstate their most likely outcome. On τ-bench, faster local decisions cut median episode time but lower task success as errors accumulate. A bigger model that makes fewer errors can be cheaper in a closed loop even if each call is slower.
The small end: 0.5B to 0.8B
Kev-0.5B is the original prototype: a 9.3M-parameter LoRA plus pointer head on Qwen2.5-0.5B, trained on a laptop in September 2026. It reports 0.799 accuracy on a held-out split of its six training sources, with ECE 0.065 raw and 0.031 after temperature T=1.47. The card labels it a superseded prototype not intended for production decisions affecting people, and its training used at most 384 state tokens. It exists to show the mechanism and reproduce the architecture, not to run a business.
Kev-0.6B is the speed play. It runs 0.12 s per five-question request on Apple Silicon against 0.33 s for Kev-0.8B, and it was trained on the decision-v7 suite in about 12 minutes on one H100. The cost is accuracy and safety: held-out policy reasoning is 6–11%, 11% of out-of-domain answers are confident errors at 0.9 or above, and the card says to use Kev-4B for accuracy. If your gate can tolerate a 0.62 transfer score and you need sub-second local decisions, it fits. If it cannot, the savings are not real.
Laya Coding Decider v2 is a specialist at roughly the 421M Laya size. It fine-tunes Laya on about 9,000 SWE-Gym-derived coding rows and reports 0.833 overall accuracy, 0.907 relevance against a majority-drop baseline near 0.64, ECE 0.147, Brier 0.086, and score MAE 0.411 on 1,150 held-out rows. The p50 is 126.4 ms, but the card does not document the hardware. This is the pattern for small specialists: a narrow task, a large gain over the baseline, and self-reported evidence with no third-party replication.
The 3B–4B workhorses
The middle is where most production pilots land. Train-Jev 4B runs on any 24 GB GPU at 43 ms p50 on an H20, scores 0.647 PK accuracy with ECE 0.076, and reports a Decision Index of 40.4 against Kev-4B's 38.4 at the same size. Its Kev transfer-v4 score is 0.779, below Kev-4B's 0.817, so it wins on one suite and loses on another. The card discloses two limits that matter: only the first 16 choice labels A–P are trained, and the results come from a single seed.
Standard One 3B is a merged BF16 3B checkpoint on Ministral-3-3B-Instruct-2512-BF16, served through SGLang 0.5.20 plus jev-adapter. Its v2.2 card reports 100.00% on JevBench public easy (48 items), 93.06% on standard (72), and 45.95% on hard (111), plus 79.97% on many-option questions with 53–151 options, up from 48.00% in v2.1. The training mixture covers nine languages. These are the publisher's own served measurements, and the card does not publish a latency figure, so the accuracy is more useful than the speed claim.
Kev-4B remains the reference at this tier: 24 GB GPU or a 32 GB Apple Silicon Mac, 1–255 options, 0.838 transfer-v4, and a single temperature T=2.41. It is not the newest model, but it is the one with the clearest published transfer numbers and a permissive licence chain.
The 9B–27B tier
The Decision Index results give a within-family view. AutoTrust's JEV-9B scores 43.14, JEV-27B scores 53.30, and JEV-Gemma4-26B-A4B scores 58.05 on Decision Index 0.2.1, each over 150,317 scoreable requests with zero errors. The 9B-to-27B jump is about 10 points; the 26B mixture-of-experts model beats the dense 27B by about 5. The runs used one B200 with 183 GB and a sequential median latency of 77 ms.
Train-Jev 27B is the independent check on that tier. It reports 0.851 Kev transfer-v4 against 0.852 for JEV-27B-VL, ECE 0.053 against 0.097, NLL 0.411 against 0.460, and Decision Index 54.6 against 54.1. It needs an 80 GB GPU and runs at 82 ms p50 on an H20. The card says the whole table is self-reported, calibration holds only in distribution, only labels A–P are trained, and the result is one seed. A statistical tie with better calibration is a real reason to choose it; it is not a reason to claim a new state of the art.
The 27B tier is also where context and option count start to matter more than the headline. A 24-slot head caps the choice labels; a 262,144-token context in the official results card is not the same as a validated 8,192-token training regime. If your workload has 40 options or 30,000-token traces, test those before comparing accuracy on short multiple-choice items.
The 35B ceiling
AutoTrust JEV-35B is the top of the open spectrum in the catalog. It is a Qwen3.5-35B-A3B mixture-of-experts model with about 3B active parameters, reads up to 256 options in one pass, and reports Decision Index 0.3 public 59.73 against JEV-27B-VL's 53.64. It scores 51.3% on ERQA in one pass with zero robot training, level with PhysBrain 1.5's 52.8. The caveats are substantial: the numbers are AutoTrust's own runs with its own kit, ERQA has 400 questions and a 95% interval of about ±5 points, and the 66.5 GiB bf16 checkpoint needs a 72 GB download and a large GPU. The repo ships a LICENSE file whose terms the card does not name.
The 35B does not make the 27B obsolete. It wins on the Decision Index public suite, but the index is a public suite, not your task. It costs more memory, and its accuracy advantage is within the range where a calibration difference or a question-design change can flip the decision. Use it when the extra capacity is doing measurable work on your rows, not because it is the largest number in the table.
Pick by constraint, not rank
| Constraint | Model | Why |
|---|---|---|
| 4 GB GPU, short states | Kev-0.8B or Kev-0.6B | Fast and small, but transfer accuracy is 0.62–0.70 |
| Coding decisions, low latency | Laya Coding Decider v2 | 0.833 and 0.907 relevance on coding rows, 126.4 ms p50 |
| 24 GB GPU, general decisions | Train-Jev 4B or Kev-4B | 43 ms p50, 0.647 PK / 0.838 transfer, different trade-offs |
| 3B with many options | Standard One 3B | 79.97% on 53–151 options; publisher-run numbers |
| 80 GB GPU, accuracy first | Train-Jev 27B | 0.851 transfer, ECE 0.053, 82 ms p50 |
| 35B with 256 options | JEV-35B | 59.73 Decision Index 0.3, 66.5 GiB bf16 |
The honest summary is that size helps within a family, specialists beat generalists on their own task, and the published numbers rarely line up across families. Choosing an open decision model turns the constraints into a sequence, the open decision-model roundup covers the wider field, and the Decision Index panel explains what the leaderboard ranks. Read decision model and decision index before comparing scores, and treat calibration as the axis that decides whether a confidence gate survives the size change.