Jev directory
CommunityJev Models & Reimplementations
Decider 0.8B
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.
- Reading time
- 2 min
- Published by
- Community
- Author
- Mapika
- Last verified
- Oct 2026
Highlights
- 1.4 GB bf16 checkpoint on Qwen/Qwen3.5-0.8B-Base; one epoch over 1.47M examples and 455M tokens on one GH200.
- In-task accuracy 0.776 with ECE 0.032 across 69 tasks; held-out 0.707 with ECE 0.096 across 24 tasks.
- Within 1 to 4 points of decider-2b on routing, classification, yes/no judgments and JSON lookups, 1.5x faster.
- Gives up knowledge, not format: TruthfulQA 0.41 vs 0.55, OpenBookQA 0.67 vs 0.81 and ARC 0.74 vs 0.86.
- Same interface as the family: decider.infer.Decider and POST /v1/systemone in TypeSafe Jev's format.
Quickstart
python
from decider.infer import Decider
d = Decider("Mapika/decider-0.8b")
d.system_one(
{"ticket": "I was charged twice for order A-104. Please refund the duplicate."},
{"team": {"type": "choice", "instructions": "Which team should handle this?",
"criteria": {"billing": "Charges, invoices, refunds", "technical": "Bugs, outages", "other": None}},
"refund_requested": {"type": "noul", "instructions": "Does the customer ask for a refund?"}})Watch out
Card reviewed 2026-10-17; all comparisons are self-reported on the card, which says it is a weaker game player (CliffWalking fails, held-out Freeway scores 0) and that what it gives up against the 2B is knowledge, not the decision format. The numbers come from one training run.
Questions about this entry
Short answers drawn from the linked project or page.
The smallest decider: a language model that does not generate text. It reads a state (a string or any JSON value) and typed questions and returns a probability distribution for every question from one forward pass, with no decoding or parsing.
It uses the same wire format as TypeSafe Jev's System One API, POST /v1/systemone, and shares the family interface decider.infer.Decider. The card calls the series a reproduction of the System One model class.
Choice with 2 to 255 options, optionally described; Score with 2 to 10 described levels; and Noul, the probability of yes. Output stays inside the options you define, and no output is generated outside them.
On the same 93-task protocol: in-task 0.776 versus 0.809 and held-out 0.707 versus 0.739, with ECE 0.032 versus 0.030 in-task. Routing, classification, yes/no and JSON lookups stay within one to four points.
Yes: the regression set runs about 1.5 times faster than on the 2B, and the bf16 weights are 1.4 GB. The card lists the 2B at 4 ms per request with CUDA graphs on one GPU.
Knowledge, not format: TruthfulQA 0.41 versus the 2B's 0.55, OpenBookQA 0.67 versus 0.81, HellaSwag 0.76 versus 0.88, ARC 0.74 versus 0.86, and wide label sets needing fine distinctions like TREC-fine.
One epoch of scripts/train.sh full from Qwen/Qwen3.5-0.8B-Base over the full mixture: 1.47M examples, 455M tokens, 4.5 hours on one GH200, with the same recipe as the 2B. Apache-2.0, and six sibling sizes share the interface.
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
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 LoRA adapter for the Standard One decision model: merge it with the supplied recipe and serve it, and it scores choice, noul and score labels over a scenario with calibrated probabilities via POST /v1/systemone, never generating text.
Jev Models & Reimplementations#community#open-models#local
Community
Nox-4B, the 4B model in vLLM Semantic Router's Decision 1.0 family: give it a state, questions and possible answers, and it returns typed decisions and probabilities with labels defined at runtime. The card reports 73.09% weighted accuracy across 3,766 decisions.
Jev Models & Reimplementations#community#open-models#local
Community
The original 0.5B Kev prototype: a 9.3M-parameter LoRA and pointer head on Qwen2.5-0.5B that reproduced the architecture Archer Hume inferred for Jev. It scored 0.799 on a held-out split of its six training sources and is superseded by Kev-0.8B, 4B and 9B.
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.