JEV-27B-VL
by autotrust
An open 27B decision model that returns a calibrated probability for every option, over text and images, in one forward pass
autotrust/JEV-27B-VLDeploy JEV-27B-VL
Single-tenantMixpeek has no managed extractor for this model. On a single-tenant deployment you upload the weights and a custom plugin serves them next to the rest of your pipeline.
Overview
JEV-27B-VL is an open vision-language decision model from autotrust, released in September 2026 under Apache-2.0 as a LoRA adapter on Qwen3.8-27B. Ask it a typed question (yes or no, pick one of up to 256 options, or a 0 to 5 rating) about text, images or both, and it returns a probability for every option in a single forward pass.
The card's headline result is zero-shot short-video recommendation: from the covers of a user's last five videos and one candidate cover, it ranks the next video as well as collaborative filtering trained on 59,045 users' watch histories (AUC 0.727 against 0.728), with no interaction data, so a new video can be ranked the moment it is uploaded.
Architecture
Qwen3.8-27B with a LoRA adapter. System 1 reads the prompt and the options and returns a calibrated probability for each option from one forward pass. System 2 is the unmodified base model, generating text and optional step-by-step reasoning, with image input. Served with vLLM; the card's serve_decide.py adds POST /v1/decide taking {kind, state, question, options}.
Mixpeek SDK Integration
# Decision models score items; they do not embed them. Index the content in
# Mixpeek as usual, then ask the model about the candidates a retriever returns.
import requests
candidates = requests.post(
"https://api.mixpeek.com/v1/retrievers/ret_your_retriever/execute",
headers={"Authorization": "Bearer API_KEY", "X-Namespace": "ns_your_namespace"},
json={"inputs": {"query": "trail running shoes for wet rock"}},
).json()["documents"]Capabilities
- Typed decisions over text and images: yes/no, pick one of 2 to 256 options, or rate 0 to 5
- A calibrated probability for every option in one forward pass
- Prompts up to 256K tokens
- The same weights also generate text and reasoning (System 2) with image input
- Served over HTTP from vLLM with POST /v1/decide
Use Cases on Mixpeek
Benchmarks
| Dataset | Metric | Score | Source |
|---|---|---|---|
| MicroLens-100k, next-video ranking from covers (zero-shot) | AUC | 0.727 | Model card: autotrust/JEV-27B-VL (self-reported; item-based collaborative filtering 0.728 using 59,045 users' histories) |
| MicroLens-100k, same task | HR@5 | 0.590 | Model card (self-reported; collaborative filtering 0.490) |
| CLINC150, 150 intents as options (zero-shot) | Accuracy | 93.8% | Model card (self-reported, with a one-line description per option; 89.5% with names alone) |
| Plan-RewardBench (1,171 trajectory pairs) | Macro accuracy | 73.2% | Model card (self-reported; GPT-5 68.5 in the paper's table) |
Performance
The card reports about 2.6 s to score 20 candidate covers on one GPU, and 46 to 300 ms per decision in its game and robotics demos. It needs a GPU large enough for a 27B model. We have not measured it.
Common Pipeline Companions
Frequently Asked Questions
What is a decision model?
A model that answers a typed question, such as yes or no or which of these options, with a probability for each answer. That makes its output easy to threshold, rank and audit. JEV-27B-VL does this over text and images.
Can JEV-27B-VL recommend videos without watch history?
On the card's MicroLens-100k test it ranked each user's next video from cover images alone with AUC 0.727, the same as item-based collaborative filtering trained on 59,045 users' histories, and a higher top-5 hit rate (0.590 against 0.490).
Can I use JEV-27B-VL commercially?
Yes. It is released under Apache-2.0.
How do decision models fit into retrieval with Mixpeek?
As a step after search: a retriever returns candidates and a decision model filters, classifies or reranks them with a probability per item. Mixpeek's built-in decision stages call TypeSafe AI's hosted Jev, which reads text only; to use JEV-27B-VL's image decisions, call your own deployment on the candidates a retriever returns, as in the example on this page.
Specification
Research Paper
autotrust/JEV-27B-VL model card
arxiv.orgBuild a pipeline with JEV-27B-VL
Add this model to a processing pipeline alongside other extractors. Combine with retrieval stages for end-to-end search.
Run it on your own data, free