Jev directory
CommunityJev Models & Reimplementations
Kev-9B
Jared Palmer's 9B Kev decision model: a rank-16 LoRA and pointer head on a frozen Qwen3.5-9B-Base that answers typed choice, score and noul questions with calibrated probabilities in one forward pass, fitting one 24 GB-class GPU.
- Reading time
- 3 min
- Published by
- Community
- Author
- Jared Palmer
- Last verified
- Oct 2026
Used inJev for Software & DevTools
- Platforms
- API
- Capabilities
- Open sourceSelf-hostedRuns locally
Highlights
- v2 is a rank-16 LoRA (45.4M parameters) plus pointer head on a frozen Qwen3.5-9B-Base, served with the adapter merged.
- Locked-test accuracy: 0.852 on transfer-v4, 0.900 on CFPB documents, 0.834 on hard skills and 0.791 on devtools.
- Serves states up to 65,536 tokens plus at least 8,192 tokens per question; the validated context length is 8,192.
- Calibration is one temperature, T = 2.19, stored in head.pt and applied at load time.
- Apache-2.0 adapter and head over an Apache-2.0 base; a self-hosted System One replacement on one 24 GB-class GPU.
Watch out
Card reviewed 2026-10-17; the locked-test scores are self-reported on the card and each partition is read once, the fetched card text truncates its out-of-scope section, and the card itself rules out fully automated decisions with legal, medical, financial or employment consequences without human review.
Questions about this entry
Short answers drawn from the linked project or page.
A decision model that reads one document (the state) and typed questions about it, then returns a calibrated probability distribution over the options supplied with each question in a single forward pass. It does not generate text and targets classification, routing, triage and policy checks.
It implements TypeSafe's public System One API (POST /v1/systemone), so the TypeSafe SDK works against it unchanged, and the card positions it as a self-hosted drop-in replacement for a System One endpoint. It is not affiliated with TypeSafe.
A frozen Qwen3.5-9B-Base backbone with 32 layers, 24 using Gated DeltaNet linear attention and 8 full attention, hidden size 4,096; a rank-16 LoRA of 45.4M parameters on its projections; and a pointer head whose softmax produces the probabilities.
The card says it fits one 24 GB-class GPU, and that it is a starting point for fine-tuning via kev.train --init_from jaredpalmer/kev-9b. States up to 65,536 tokens are served plus at least 8,192 per question, but the validated context length is 8,192 and training used at most 7,552.
They are the card's own once-read results: transfer-v4 accuracy 0.852 with Brier 0.199, breadth-v1 0.698, hard-v1 0.834, devtools-v1 0.791 and documents-v1 0.900. The card says this checkpoint is included in Kev 1.0, released 2026-09-30.
Same recipe at a different scale: rank-16 LoRA plus pointer head on a Qwen3.5 base, with v1 being the same recipe before the documents and skills stage. Kev-4B fits a 24 GB GPU or a 32 GB Mac; Kev-27B is a full-weight fine-tune at 51.3 GB bf16.
Text generation, chat, summarisation and open-ended question answering, because the model only scores the options it is given. The card also rules out fully automated decisions with legal, medical, financial or employment consequences without human review, and questions needing facts outside the state.
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
Jared Palmer's 27B Kev decision model: a full-weight fine-tune of Qwen3.8-27B averaged with v1, plus a pointer head that returns calibrated probabilities for typed choice, score and noul questions in one forward pass. Card-reported transfer-v4 accuracy is 88.87%.
Jev Models & Reimplementations#community#open-models#local
Community
The smallest Kev: an 11.3M-parameter LoRA and pointer head on Qwen3.5-0.8B-Base that answers typed choice, score and noul questions with calibrated probabilities in one forward pass, running on a laptop or a 4 GB GPU.
Jev Models & Reimplementations#community#open-models#local
Community
Jared Palmer's Kev-4B: a 33.8M-parameter rank-16 LoRA and pointer head on Qwen3.5-4B-Base that returns calibrated probabilities for typed choice, score and noul questions in one forward pass, sized for one 24 GB GPU or a 32 GB Apple Silicon Mac.
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
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
The smallest decider from Mapika: a 1.4 GB bf16 Qwen3.5-0.8B-Base fine-tune that returns calibrated probabilities for typed Choice, Score and Noul questions in one forward pass and runs 1.5x faster than decider-2b, within one to four points on routing and lookups.
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.