Jev directory
CommunityJev Models & Reimplementations
Kev-8B
Jared Palmer's Qwen3-generation Kev-8B: a rank-16 LoRA and pointer head on Qwen3-8B-Base that answers typed questions with probabilities in one forward pass. The card keeps it as the fast Apple Silicon option while Kev-9B leads on accuracy, 0.837 to 0.780.
- 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/Qwen3-8B-Base; it serves /v1/systemone with no text generation.
- Card-reported in-distribution accuracy 0.869 with raw ECE 0.061; transfer-v4 development 0.796 with Brier 0.337.
- Locked test read once: 0.870 in-distribution and 0.780 out of domain, with 7.6% confident errors out of domain.
- Kept as the fast Apple Silicon option because its attention-only backbone runs the packed forward at full speed on MPS.
- Out-of-domain probabilities are usable but not calibrated (raw ECE 0.128); 8B fp32 needs about 33 GB, bf16 about 17 GB.
Watch out
Card reviewed 2026-10-17; the scores are the authors' own, and the card itself says held-out policy reasoning is far from Jev, that temperature fitted in domain does not transfer, and that Kev-9B supersedes this checkpoint for accuracy and calibration.
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/Qwen3-8B-Base that takes one state and typed questions and returns a probability distribution per question in one forward pass, serving the /v1/systemone contract.
The card keeps it as the fast option on Apple Silicon, because its attention-only backbone runs the packed forward at full speed on MPS. For accuracy and calibration the card points to Kev-9B instead.
Decision-v4 development 0.869 with raw ECE 0.061 and transfer-v4 development 0.796 with Brier 0.337. The locked test, read once, scored 0.870 in-distribution and 0.780 out of domain with 7.6% confident errors.
On the same frozen items in its comparison table it is close in-distribution (0.863 versus Jev's 0.845 on decision-v4 dev) but clearly behind out of domain (0.796 versus 0.857) with more confident errors (9.9% versus 3.7%). Jev's training data is unknown.
8B in fp32 needs about 33 GB of memory and does not fit a 32 GB Mac; setting KEV_DTYPE=bf16 cuts it to about 17 GB. Training took about 70 minutes on one H100 according to the card.
Capacity dominates out of domain (0.6B to 4B is large, 4B to 8B small), fine-tuning erodes base capability unless the learning rate is lowered, and more public data raised in-distribution accuracy while lowering transfer at 4B.
Held-out policy reasoning is far from Jev, product-shaped questions with no training analogue are not guaranteed, out-of-domain probabilities are usable but uncalibrated (raw ECE 0.128), and a temperature fitted in domain does not transfer.
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.