Kev, Laya, and d1 get grouped together because all three answer a state plus typed questions with probabilities instead of prose. That grouping hides the important differences. Kev is a family of pointer-head checkpoints over frozen Qwen backbones. Laya is a family of encoder assessors from Convai Innovations. d1 is Liquid AI's API model, known here through a vendor benchmark claim. The architecture decides what hardware you need; the evidence decides how much of the published number to believe.
The architecture each family chose
The Kev line keeps a decoder backbone frozen and trains a small adapter plus a pointer head. Kev-0.8B adds an 11.3M-parameter rank-16 LoRA and a pointer head to Qwen3.5-0.8B-Base. Kev-4B does the same with a 33.8M-parameter adapter on Qwen3.5-4B-Base. The model reads the state once, gives each question its own branch, and scores option boundary tokens against a decide token. It supports 1–255 options and serves the same POST /v1/systemone contract as Jev.
Laya takes the encoder route. Laya ships three checkpoints: a 421M ModernBERT-large English model, a 322M mmBERT-base multilingual model, and a 421M typed-decisions checkpoint, with a router picking per request. It is trained with RLCD, reinforcement learning against strictly proper scoring rules, and answers in one forward pass. The card reports 33 ms for one question and 7.2 ms per question batched on a T4. The trade-off is context: the English checkpoint starts at 512 tokens, the multilingual model is documented at 1,024 and can be pushed to 8,192, and quality degrades past roughly 4,000 tokens.
d1 is the outlier. GIGAZINE's report describes Liquid AI's model as the first to outperform Jev on Hugging Face's Decision Index, available through Liquid's API and free then. That is a vendor run reported by a news site, not an independent rerun, and the entry gives no weight format, parameter count, context limit, or license. Treat d1 as an API-first option with a strong marketing claim, not as an open checkpoint you can audit.
What the numbers actually measure
The three families publish different numbers on different populations, which makes a single accuracy column misleading.
| Family | Published result | What it is measured on | Main caveat |
|---|---|---|---|
| Kev-0.8B | transfer-v4 0.697, Brier 0.397; documents-v1 0.851; hard-v1 0.665 | The author's held-out task families | Card-reported; out of domain it is markedly weaker than Kev-4B |
| Kev-4B | transfer-v4 0.838, Brier 0.224; hard-v1 0.803; devtools-v1 0.756 | Author's held-out task families | Card-reported; 8,192 tokens validated, 65,536 accepted |
| Laya | 0.766 on 2,000 typed decisions; base checkpoints near chance at 0.36 and 0.35 | Fine-tuned typed-decisions checkpoint | Different prompts and sample sizes; raw calibration weak |
| Laya Coding Decider v2 | 0.833 overall, 0.907 relevance, ECE 0.147, p50 126.4 ms | 1,150 held-out coding rows from SWE-Gym | Self-reported; latency hardware undocumented |
| d1 | First to beat Jev on the Decision Index, per Liquid | Liquid's own Decision Index run | Vendor claim via press; no independent replication |
The Kev numbers are transfer results, not JevBench results. The 0.766 belongs to Laya's fine-tuned typed-decisions checkpoint; the base English and multilingual models score 0.36 and 0.35 against a 0.318 baseline.
The independent comparison
The cleanest external comparison is Jev vs Laya, which runs hosted Jev 1.13 against Laya 421M on MLX over 1,470 synthetic items across eight tasks. Across 3,386 judgments, Jev scores 92.9% correct and Laya 65.3%; by question type, Jev reaches 99.3 Choice, 98.6 Noul, and 76.1 Score, against Laya's 68.3, 71.3, and 51.5. Jev wins all 19 questions with paired confidence intervals excluding zero.
The latency picture is more nuanced. On one short question, Laya's p50 is 42 ms against Jev's 136 ms, including about 50 ms of network. With 50 questions on one state in one request, Jev answers in 170 ms against Laya's 1,002 ms. It is the authors' own synthetic benchmark, and only 421M Laya checkpoints were tested, so treat it as careful development work, not independent confirmation.
A second independent study, the Laya typed-probabilistic assessor reproduction, reruns the checkpoint and gets 0.767 against the card's 0.766. It finds the model uniformly under-confident: the signed gap is -0.214, and a fitted temperature of 0.469 cuts held-out ECE from 0.204 to 0.037. That is opposite to what the card anticipates, and it changes which way a confidence gate should lean.
Where each family is strongest
Kev's strongest territory is tool and action routing, where the state is a schema-heavy agent context. Soro-Kev-4B is a pointer-head descendant that reports 89.00% zero-shot tool routing, 94.50% on external human holdouts, and 75.76% calibrated abstention on BFCL irrelevance slices. It fits in about 4,048 MB at NF4, or 8,200 MB at fp16, and reports a 413 ms median prefill on an RTX 4050 Laptop GPU. Those are self-reported, and training states were capped at 4,096 tokens even though the card advertises 32,768.
OpenJev shows the other side of the Kev-style bet. Its Qwen3.5 NLI cross-encoder scores 0.814 on JevBench v1.2 public items against Jev 1.13's 0.866, and 0.622 on the hard tier against 0.730. Option reordering changes zero labels over 231 items, with a largest probability shift of 1.8e-7. But the author measured 0.600 accuracy as a shell-command guard, catching 37% of deny-worthy commands, and a single adversarial line dropped accuracy from 0.833 to 0.467. A pointer head gives you control and a small memory footprint; it does not give you a security boundary.
Laya's strongest territory is multilingual and CPU-adjacent work. Laya-MLX runs the checkpoints natively on Apple silicon and reports a 13.4 ms median for a short English decision and 7.4 ms on the multilingual checkpoint. It matched upstream's selected answer on 63 of 63 validation questions in FP32 and FP16, which is fidelity on fixtures rather than accuracy on your task. The coding fine-tune is the clearest specialist: about 9,000 SWE-Gym-derived rows, 0.833 overall accuracy, 0.907 relevance against a majority-drop baseline near 0.64, and a 126.4 ms p50. The card does not name the hardware.
d1's strongest territory is the one the other two cannot serve: a hosted API with no weights to download. You cannot inspect the checkpoint, rerun the training mixture, or verify the Decision Index claim outside Liquid's own execution. Ask for per-axis numbers and a held-out set before you move a workflow onto it.
How to choose
Start with the constraint that cannot move.
- Four gigabytes of VRAM or a laptop: Kev-0.8B, the small pointer-head option. Its 0.697 transfer score is honest about the capacity limit.
- One 24 GB GPU or a 32 GB Mac: Kev-4B, with 0.838 transfer accuracy and 1–255 options.
- Multilingual or CPU-only: Laya, with the encoder checkpoints and the MLX path. Check the checkpoint name and the 1,024-token default first.
- Coding decisions: Laya Coding Decider v2, but its 0.833 and 126.4 ms are one author's held-out split.
- No operations budget: d1, the API route. Treat the Decision Index headline as a claim to test.
What to verify before committing
Three checks separate a pilot from a demo. Identify the benchmark: transfer-v4, JevBench, and a private coding split are not interchangeable. Check calibration direction: Laya is under-confident by 0.214 before fitting, and Kev publishes one temperature per checkpoint. Check the license chain: Kev and Laya are Apache-2.0, while d1's API terms are separate from open weights.
The open decision-model roundup covers the wider field, and choosing an open decision model turns these constraints into a buyer's sequence. For the measurement side, measuring Jev explains why one accuracy number rarely survives production. The useful unit is not "which family wins" but "which family's evidence matches the decision," with calibration and open weights treated as requirements.
