Jev directory
CommunityJev Models & Reimplementations
Kev-0.6B
The small Qwen3-generation Kev: a rank-16 LoRA and pointer head on Qwen3-0.6B-Base that answers typed questions in one forward pass and stays the fastest small Kev on Apple Silicon at 0.12 s per five-question request, though out-of-domain transfer sits near 0.62.
- Reading time
- 2 min
- Published by
- Community
- Author
- Jared Palmer
- Last verified
- Oct 2026
Used inJev for Software & DevTools
- Platforms
- API
- Capabilities
- Open sourceSelf-hostedRuns locally
Highlights
- LoRA r=16 plus pointer head on Qwen/Qwen3-0.6B-Base; decision-v4 development accuracy 0.801 with raw ECE 0.086.
- Transfer accuracy 0.620 out of domain, flat at about 0.60 across every hyperparameter the authors tried.
- Locked test read once: 0.808 in-distribution (ECE 0.089) and 0.642 out of domain (ECE 0.128).
- Fastest small Kev on Apple Silicon at 0.12 s per five-question request, against 0.33 s for Kev-0.8B.
- Documented limits: held-out policy reasoning 6-11% and 11% confident errors out of domain.
Watch out
Card reviewed 2026-10-17; the figures are the authors' own, and the card says to use Kev-4B for accuracy and this checkpoint only where memory or latency rule the 4B out, measuring on your own data. Confident out-of-domain errors run at 11%.
Questions about this entry
Short answers drawn from the linked project or page.
A previous-generation Kev decision model: LoRA r=16 plus a pointer head on Qwen/Qwen3-0.6B-Base that serves the /v1/systemone contract, returning a probability distribution per typed question in one forward pass instead of generating text.
Speed on Apple Silicon: the card says 0.12 s per five-question request versus 0.33 s for Kev-0.8B. For accuracy it points to Kev-0.8B, which scores 0.668 against 0.642 out of domain on the same locked test items.
0.620 on transfer-v4 development and 0.642 on the locked test, with the card saying transfer is flat at about 0.60 across eight hyperparameter mutations and three seeds. Capacity, not data, is the bottleneck at this size.
Training distributions: decision-v4 development 0.801 with raw ECE 0.086. Out of domain: transfer-v4 0.620 with Brier 0.536, locked-test ECE 0.128 and 7.9% confident errors, plus per-source scores like QNLI 0.85 and MMLU 0.50.
Held-out policy reasoning fails (6-11% both-siblings-correct), 3-level score questions with date arithmetic collapse to the middle level, and 11% of out-of-domain answers are confident errors at 0.9 or above. Probabilities are advisory out of domain.
On the decision-v7 suite: 10,000 public records, 896 policy minimal-pair records and 1,680 records from 60 random rule structures, two epochs, LoRA r=16 at learning rate 1e-4, about 12 minutes on one H100. No Jev outputs were used for training.
Apache-2.0, on Qwen/Qwen3-0.6B-Base. The card marks it previous generation and keeps it for speed; the locked test was read once and will not be read again for this checkpoint.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: Standard One SH (StandardThinking)
LinkedIn: What multimodal decision models actually change, and what they do not.
LinkedInJevDirectory.org
Hacker News: Decision Index 0.2.1 scores open reproductions of Jev on a frozen suite.
jackbravo4 pointsDecision Index 0.2.1 – 70 open reproductions of TypeSafe Jev's Decision ModelRead the thread on Hacker NewsHacker News: Pointer to a DiffusionGemma plus vLLM tutorial for a Jev-shaped classifier.
1 pointHow to build a Jev-style classifier with DiffusionGemma and vLLMRead the thread on Hacker NewsMore Jev resources like this
A LoRA plus 24-slot decision head over Qwen3.8-27B from Ming Xu's train-jev project: it serves TypeSafe's System One API with 0.851 Kev transfer-v4 accuracy and ECE 0.053, and needs an 80 GB GPU.
Jev Models & Reimplementations#community#open-models#local
Community
A 24 GB-GPU alternative from the train-jev project: Qwen3.5-4B plus a rank-16 LoRA and 24-slot decision head scoring 0.647 PK accuracy, 40.4 Decision Index and 43 ms p50 latency, ahead of Kev-4B at the same size.
Jev Models & Reimplementations#community#open-models#local
Community
Singularity Compute's fine-tune of Google's diffusiongemma-26B-A4B-it, a discrete-diffusion mixture-of-experts, served from an 18 GB NVFP4 checkpoint that answers choice, noul and score questions with calibrated distributions on one 48 GB GPU.
Jev Models & Reimplementations#community#open-models#local
Community
Standard Thinking's 3B decision model with a joint schema head: it scores every option of every question in a scenario in one forward pass through POST /v1/systemone, and v3 self-scores 51.53 on the Decision Index 0.3 public index, up from v2's 42.90.
Jev Models & Reimplementations#community#open-models#local
Community
Mapika's NVFP4 fine-tune of Gemma-4-31B-it answers typed Choice, Noul and Score questions with calibrated probabilities and no generated text, self-reporting a 64.27 chance-corrected public Decision Index and 39 ms median latency on a B300.
Jev Models & Reimplementations#community#open-models#local
Community
Standard Thinking's 8B decision model with a joint schema head that reads the scenario once and scores every option of every question; v4 self-scores 57.40 on the Decision Index 0.3 public index, up from v3's 48.34, with a 70.6M-parameter head.
Jev Models & Reimplementations#community#open-models#local
Community
From the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.