CommunityJev Models & Reimplementations
Jev-alpha-26B-A4B
chenz53's standalone decision model derived from Google Gemma 4 26B A4B IT: full merged BF16 weights, a fitted temperature of 1.736507, and 79.50% equal-type accuracy over 1,100 final questions.
- Published by
- Community
- Author
- chenz53
- Last verified
- Oct 2026
- Languages
- Python
- Capabilities
- Open sourceSelf-hostedRuns locally
Highlights
- Merged BF16 weights derived from google/gemma-4-26B-A4B-it; no LoRA adapter is needed at inference.
- Equal-type final accuracy 79.50% on 1,100 questions, weighting noul, choice, and score equally.
- Fitted temperature 1.736507 from 300 separate calibration questions; BF16 merging changes numerical behavior.
- Reads final-position logits directly with label ids instead of calling generate; 2-26 options per question and a 4,096-token context limit.
- Trained on 8,400 questions for two epochs with LoRA rank 8, alpha 16, and dropout zero; seed 42, KL weight 0.1, checkpoint 2100.
Watch out
Apache-2.0 for the model, with Google DeepMind attribution in NOTICE and BASE_MODEL_CARD.md; SciQ training data carries CC-BY-NC-3.0 terms. Final controls come from the same source families as training, the Jev comparison is descriptive rather than a paired benchmark, and the card says the release does not establish Jev-equivalent speed, cost, or preservation of all foundation-model abilities.
Questions about this entry
Short answers drawn from the linked project or page.
Google Gemma 4 26B A4B IT, at base revision 4d7ae4984b7db7de8f8457170b3f1a419ee76d52. It retains the upstream Gemma4ForConditionalGeneration architecture and tokenizer, and the repo ships full merged BF16 weights with no adapter needed at inference.
79.50% equal-type final accuracy on 1,100 questions, with each source contributing 100 and equal weight on noul, choice, and score. Per-dataset accuracy runs from 53.00% on sst5 and 61.00% on yelp5 to 100% on dbpedia14 and sciq.
With Transformers 5.16.1, PyTorch 2.8.0, BF16, SDPA, and grouped_mm, as tested on an H200. The example reads final-position logits directly and never calls generate; a generic text-generation pipeline does not return this decision schema.
1.736507, fitted on 300 separate calibration questions. The card warns the merged BF16 model behaves differently from the earlier adapter or three-seed mean, so it says to use the release results and this temperature rather than those.
The card says no: it is descriptive, not a paired benchmark or a claim of parity. Jev values are third-party measurements published by Kev, question samples, prompts, and option construction differ, and Jev's training data is unknown.
Final controls come from the same source families as training, BoolQ controls are unused cleaned training records, the When2Call public subset covers only clarification and cannot-answer targets, and new authored questions and action labels lack independent human review.
The model is Apache-2.0 through Gemma 4, with LICENSE, NOTICE, and the upstream base-model card included. SciQ carries CC-BY-NC-3.0 terms and other sources have their own terms; the model licence does not grant rights to the source datasets.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
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 NewsHacker News: A one-token logprob wrapper that also classifies webcam frames.
allanrbo151 pointsA single function Jev-like wrapper for LLMs, including vision modelsRead the thread on Hacker NewsRelated terms
Glossary definitions related to this entry.
More Jev resources like this
AutoTrust's combined System 1 and System 2 bundle on Gemma-4-26B-A4B-it: Decision Index 0.2.1 of 58.05, a 24-slot head with noul, choice and score, and the base chat model left bit-identical.
Jev Models & Reimplementations#community#open-models#local
Community
alibiserikbay's Apache-2.0 Jev alternative on Qwen3.5-4B with a merged rank-16 LoRA: one forward pass returns a probability per option, 0.784 on JevBench's public hard tier at about 13 ms per decision.
Jev Models & Reimplementations#community#open-models#local
Community
The 9B JevK5: the same v0.3 data on Qwen3.5-9B, retrained for 1.5 epochs by a pre-registered sweep, reaching index proxy 0.794 and a 0.859 hand-written hard set, Apache-2.0.
Jev Models & Reimplementations#community#open-models#local
Community
A 0.6B non-autoregressive decision model on Qwen3-0.6B that returns calibrated Choice, Score and Noul distributions in one forward pass, scoring 95.8% on its 22,110-item held-out test split.
Jev Models & Reimplementations#community#open-models#local
Community
An untrained readout: stock Gemma-4-31B weights plus a decider config that reads option letters at the answer slot, scored #2 of 70 on Decision Index v0.2.1 at 57.33 against Jev 1.13.0's 57.91.
Jev Models & Reimplementations#community#open-models#local
Community
chaoliangUNSW's v1 Jev-Style decision model in GGUF from 1.3 GB: calibrated choice, bool and score answers on Qwen3.5-2B-Base, 82.3% on five held-out task families with ECE 0.017.
Jev Models & Reimplementations#community#open-models#local
Community
From the Jev guides
Original write-ups on what this entry is for.
Training recipes, looped architectures, and language ports: what the open side of the typed-decision class shipped this month, and how to pick by runtime and calibration instead of the leaderboard.
Jev Models & Reimplementations#open-models#training#local
Read article
Taking a decision checkpoint out of Python: what the GGUF and browser WebGPU exports preserve, what they change, and the parity evidence the model cards publish.
Jev Tools & CLIs#open-models#local#onnx
Read article
Interface, size, license, claims — in that order. A selection guide across the system-one catalog, from a 706K form scorer to 27B Jev students.
Jev Practices & Patterns#open-models#evaluation#benchmarks
Read article
From the Jev community
Posts from builders shipping with Jev right now.