Jev directory
CommunityJev Models & Reimplementations
Decider 4B
Mapika's 4B decider: a Qwen3.5-4B-Base fine-tune that returns probabilities over supplied options for typed questions in one forward pass, with no decoding. Version 2.1 adds a rank-64 LoRA stage and reports a JevBench public hard tier of 0.649.
- Reading time
- 3 min
- Published by
- Community
- Author
- Mapika
- Last verified
- Oct 2026
Used inJev for Software & DevTools
- Capabilities
- Open sourceSelf-hostedRuns locally
Highlights
- v2.1 is v2 plus a rank-64 LoRA stage trained for 2 epochs on 29,325 harder decision rows, merged into the weights.
- Base is Qwen/Qwen3.5-4B-Base: 4.2B parameters, 32 layers, 24 with linear attention and 8 with full attention; 8.4 GB bf16.
- Card-reported JevBench public hard tier 0.649, held-out generated decision families 0.556 and regression set 0.831 / 0.784.
- One temperature per answer type: 1.110 choice, 1.560 noul and 1.287 score, over a base temperature of 1.099.
- No reinforcement-learning stage; v2 and v1 remain on the Hub under their own tags for comparison.
Watch out
Card reviewed 2026-10-17; all comparisons are self-reported, and the card lists regressions: v2.1 is less well calibrated than v2 on hard items, still answers one of the two form-filling cases of issue #9 wrongly, and is worse than v1 on BabyAI-GoTo and greedy bag-draw play. The fetched card text truncates its comparison table.
Questions about this entry
Short answers drawn from the linked project or page.
An open decision model from Mapika that reads a state and one or more typed questions with explicit option lists and returns a probability distribution for every question from one forward pass. There is no decoding, no parsing and no output outside your options.
The card calls it an open reproduction of the System One model class, TypeSafe AI's Jev. It is called from software, not chatted with, and it returns probability distributions rather than generated text.
Replay rows are now trained toward v1's own answer distribution with a KL loss instead of hard labels, which lowered the fitted temperature from 1.935 to 1.099, and temperatures are now set per answer type. Rows, LoRA rank and epochs are unchanged.
The base Qwen3.5-4B-Base holds 4.2B parameters across 32 layers, 24 with gated delta-net linear attention and 8 with full attention, with hidden size 2,560. The repository holds 8.4 GB bf16 weights; the card does not state hardware requirements or latency.
Its own comparisons against v2 and v1: JevBench public hard tier 0.649 versus 0.676 and 0.550, held-out generated decision families 0.556 versus 0.560 and 0.469, live browser tasks 93.2% versus 88.1% and 90.9%, all self-reported.
It is less well calibrated than v2 on hard items, still answers one of the two form-filling cases of issue #9 wrongly, is worse than v1 on BabyAI-GoTo and greedy bag-draw play, and trails decider-35b-a3b by 2.4 and 2.6 points on the regression set.
Mapika. It is a fine-tune of Qwen/Qwen3.5-4B-Base trained in two supervised stages: a cross-entropy pass over the mixture v2 (742M tokens) and a LoRA rank-64 second stage on 29,325 rows, with no reinforcement-learning stage.
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
Cloudflare's 9B multimodal decision model on Ollama: text, JSON, or images in one non-autoregressive forward pass, Apache-2.0, and the lowest measured latency of any decision model in Cloudflare's Decision Index runs at 38.8 ms median.
Jev Models & Reimplementations#community#open-models#local
Community
Bespoke Labs' 9B decision model on Ollama, fine-tuned from Qwen3.5-9B: Apache-2.0, up to 64 typed questions per call, answers in under 100 ms on a MacBook Pro with an M5 Max, and 75.7% mean accuracy across 13 human-labelled datasets.
Jev Models & Reimplementations#community#open-models#local
Community
A 70.8M-parameter DeBERTa-v3-xsmall cross-encoder that answers typed decision questions in one batched pass with zero generated tokens, shipped with a calibration.json that the engine requires.
Jev Models & Reimplementations#community#open-models#local
Community
A pointer-network decision model from Keyush Nisar that reads one agent state document plus typed questions and returns calibrated tool-routing probabilities in a single forward pass, reporting 89.00% zero-shot tool routing.
Jev Models & Reimplementations#community#open-models#local
Community
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
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.