Jev directory
CommunityJev Models & Reimplementations
PPLX Decider v1.1 27B
Perplexity AI's 27B decision checkpoint, a fine-tune of Qwen/Qwen3.8-27B with a BF16 decision head and saved calibration temperature; the card reports a Decision Index of 61.56, up from 56.4 in v1 and above Jev by more than 3.5 points.
- Reading time
- 2 min
- Published by
- Community
- Author
- Perplexity AI
- Last verified
- Oct 2026
Used inJev for Software & DevTools
- Languages
- Python
- Capabilities
- Open sourceSelf-hosted
Highlights
- Native decision-checkpoint layout: a Qwen3_5Model backbone plus readout.safetensors containing a BF16 [255, 5120] decision head.
- Card reports Decision Index 61.56 overall, beating v1's 56.4 and Jev 1.13's 57.9; best gains in Language 69.45 and Retrieval 61.26.
- Full-attention layers use the saved noncausal attention mode; default causal inference does not reproduce the evaluated behaviour.
- Needs Python 3.12+, authenticated Hugging Face access, and a CUDA GPU with room for roughly 49 GiB of weights plus working memory.
- Images are supported: a 32x32 patch costs one visual token and max_image_pixels defaults to 2048x2048 (up to 4,096 tokens).
Quickstart
python
import sys
from pathlib import Path
from huggingface_hub import snapshot_download
checkpoint = Path(snapshot_download("perplexity-ai/pplx-decider-v1.1-27b"))
sys.path.insert(0, str(checkpoint / "source" / "src"))
from autojev.model import DecisionModel, answer
model = DecisionModel(checkpoint, device="cuda")Watch out
Card reviewed 2026-10-17; all Decision Index numbers are self-reported from the card's own panel and the repository requires authenticated Hugging Face access. The supplied DecisionModel.predict applies the saved calibration temperature, and the card warns that default causal inference, a plain lm_head or applying temperature more than once will not reproduce the evaluated setup.
Questions about this entry
Short answers drawn from the linked project or page.
An updated 27B decision model from Perplexity AI, fine-tuned from Qwen/Qwen3.8-27B. It lifts the earlier pplx-decider-v1-27b's Decision Index from 56.4 to 61.56 while using the same backbone, mostly through lifting the causal mask and training on more data.
It is a competing decision model evaluated on the same Decision Index suite: the card reports 61.56 overall, ahead of Jev 1.13's 57.9 by more than 3.5 points, though Jev still leads the Knowledge category 51.4 to 48.18.
A native decision-checkpoint layout: a Qwen3_5Model backbone plus a readout.safetensors file holding a BF16 [255, 5120] decision head. It is a separate readout, not a full-vocabulary lm_head, so serving systems expecting Qwen3_5ForConditionalGeneration need a separate export.
The card asks for Python 3.12 or newer, authenticated Hugging Face access to the private repository, and a CUDA GPU with room for roughly 49 GiB of weights plus working memory, using the pinned requirements.txt dependencies.
The card's panel reports overall 61.56 versus v1's 56.4 and Jev's 57.9, with Language at 69.45 and Retrieval at 61.26 as the largest category gains. The overall figure uses the suite's weighting, and all numbers are self-reported.
Yes; images go in the images field as file paths, PIL images or base64 data URLs. The processor preserves aspect ratio, rounds each side to a multiple of 32 px, and the text plus all images must fit within prepare's 8,192-token limit.
Default causal inference does not reproduce the evaluated behaviour: full-attention layers need the saved noncausal mode, DecisionModel.predict applies the saved calibration temperature, and when using logits directly you must apply the temperature once and normalize over valid candidates.
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
X: Standard One SH (StandardThinking)
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
A small distilled sibling of Perplexity's pplx-decider-v1.2: a Qwen3.5-4B decider that reads a state plus typed choice, noul and score questions in one pass, keeping most of the 27B model's Decision Index.
Jev Models & Reimplementations#community#open-models#multimodal
Community
A Laya fine-tune for coding decisions: roughly 9k rows from SWE-Gym teach keep/drop file relevance, priority scoring and injection detection, reporting 0.833 accuracy on 1,150 held-out rows at 126 ms p50.
Jev Models & Reimplementations#community#open-models#local
Community
Perplexity's 27B decision model rebuilt on Qwen3.8-27B with a joint request format that answers a whole request's questions in one sequence, reporting 68.09 on Decision Index 0.3 and 77.43 on vision.
Jev Models & Reimplementations#community#open-models#multimodal
Community
A Laya fine-tune from Convai Innovations on the 1,200-case typed-decisions benchmark: it reports 0.777 accuracy on the official 400-case test set, above TypeSafe Jev 1.13.0 at 0.727 and the teacher ceiling at 0.735.
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.