# JEV-27B-VL (slim)

> JEV-27B-VL is AutoTrust AI Lab's open-weight decision model, a LoRA adapter and decision head on Qwen3.8-27B that returns option probabilities for questions over text and images, trained on labels distilled from TypeSafe's Jev. It is not a TypeSafe product.

- Full: https://www.anchorterminal.com/tools/jev-27b-vl.md (~8,300 tokens) · this version ~2,330 tokens · JSON https://www.anchorterminal.com/tools/jev-27b-vl.json · canonical https://www.anchorterminal.com/tools/jev-27b-vl
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-10-21 · pricing https://www.anchorterminal.com/tools/jev-27b-vl#pricing · reviews https://www.anchorterminal.com/tools/jev-27b-vl#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**C · 59.5/100 · rank #60 of 100 open-weight models · #13 in Decision models · not agent-ready · confidence low**


Assessment: A cousin for typed decisions with probabilities over text and images, answering in one pass with optional thinking. It needs about 67 GB in BF16 on vLLM, with no AutoTrust 4-bit build or hosted API. The main risks are labels distilled from TypeSafe's Jev against its customer terms and serving through a repository script on a vLLM development build.

## Facts

- Kind: Open-weight model · vendor: AutoTrust AI Lab · category: Decision models · legal entity: not named · provenance 13/100
- Packages: huggingface `autotrust/JEV-27B-VL`
- Auth: None · pricing: Free · x402: no · licence: Apache-2.0 for AutoTrust's System 1 adapter and decision head. The base Qwen3.8-27B weights in the repository are Alibaba's, unchanged, also Apache-2.0
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights at huggingface.co/autotrust/JEV-27B-VL, read from the model card, the files and their metadata, serve.sh and serve_decide.py, the adapter and decision-head files, the licence, the training corpus's card and the runtimes' own code on 21 October 2026. Nothing was downloaded or run. The live demo at news.guru.so isn't graded
- Built on: Qwen/Qwen3.8-27B (Apache 2.0), declared in the card's metadata as an adapter. The repository carries the base weights byte for byte, plus AutoTrust's System 1 LoRA (rank 32, alpha 64, on the attention, linear-attention, MLP and lm_head projections) and a decision head that reads 24 verbaliser tokens with a bias and per-kind temperature
- Relation to TypeSafe's Jev: A different organisation. AutoTrust trained the adapter to imitate TypeSafe Jev 1.13's output distributions and reports a mean KL of about 0.017 against them on a held-out set. Its JEV-27B card disclaims any affiliation with TypeSafe AI. It shares Jev's question types (noul, choice, score) but not its API path or model names
- Model: 27.8B parameters in the base files (vision encoder and multi-token prediction head included), 64 layers in a 3 to 1 mix of Gated DeltaNet and Gated Attention. System 1 is the adapter, one forward pass per decision. System 2 is the unchanged Qwen3.8-27B, which can think first and have its answer folded in
- Files: 18 safetensors shards in BF16 (55.6 GB), adapter_vllm/adapter_model.safetensors (0.45 GB), decision_head.json, calibration.json, chat_template.jinja, configs and tokenizer, plus serve.sh and serve_decide.py, demo videos and images. No pickle files. Hugging Face lists a SHA-256 for every LFS file
- Licence: Apache 2.0 for AutoTrust's files, and the base's Apache 2.0 for Qwen's. No use policy in the repository. The training labels came from a hosted model whose customer agreement forbids distillation, which the licence doesn't address
- Runtimes: The card documents vLLM only, through serve.sh and serve_decide.py (a vLLM development build of September 2026) or plain vllm serve with --enable-lora. Released vLLM (0.30.0) and SGLang (0.5.22) apply LoRA to the qwen3_5 architecture including lm_head, and Transformers runs qwen3_5 with PEFT adapters. No GGUF of the adapter and no Ollama or LM Studio entry
- Context: 262,144 tokens native, per config.json and the card, which reports 20 of 20 single-fact decisions right at every length up to 250K. serve.sh starts at 32,768 unless MAX_MODEL_LEN is set
- Inputs and outputs: POST /v1/decide takes kind (noul, choice with 2 to 256 options, or score 0 to 5), a state of text, JSON or text mixed with images, and one question, and returns a probability for every option with the top choice. thinking off (default), auto or on adds System 2, with reasoning_effort, think_budget and return_reasoning. System 2 is also served as an ordinary OpenAI chat model
- Hardware: The card says one GPU with 80 GB or more, about 52 GB of weights plus about 65 KB of KV cache per token, and --max-num-seqs 8, since larger batches give wrong System 1 probabilities on vLLM's multimodal LoRA path
- Training data: Published, SargeDev/jev-distill-corpus-v3 (Apache 2.0), 740,957 rows from three streams with their origins and sizes, synthetic scenarios labelled by Jev 1.13, a 32B teacher's labels over public QA sets and CC0 Open-Jev rows. Text only, so image decisions are zero-shot. The base model's pretraining data isn't disclosed

- Scores: Capability pending, Reach 50, Risk 45, Agent interface 80, Documentation and upkeep 92 · total over the 7 assessed categories
- Why:  · Reach, Memory 10 of 30, the only build is BF16 at 56.0 GB with the adapter and vision encoder, 67.2 GB with a fifth added, which fits the reference… · Risk, File safety 15 of 25, safetensors only, but the documented quick start runs the repository's serve_decide.py. · Agent interface, Read as a decision model, by what it takes and returns. · Documentation and upkeep, Model card 17 of 20, everything but a knowledge cut-off.
- Sources: 19, open questions: 4, both in the full twin
- Capabilities: inference.decision, model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/jev-27b-vl.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/jev-27b-vl.svg` or a link to https://www.anchorterminal.com/tools/jev-27b-vl from a page under huggingface.co/autotrust, then `POST https://www.anchorterminal.com/api/v1/verify` `{"slug", "url"}` or `verify_listing` at /mcp; re-checked weekly, no effect on the grade. Snippets in the full twin.

## Before you call it

1. Quickest start is `hf download autotrust/JEV-27B-VL` then `bash JEV-27B-VL/serve.sh`, which serves POST /v1/decide on port 8000, but it runs the repository's serve_decide.py on a September 2026 vLLM development build
2. On a released vLLM use plain `vllm serve` with `--enable-lora --lora-modules jev-decision=JEV-27B-VL/adapter_vllm` and do the card's client-side arithmetic with decision_head.json and calibration.json
3. Keep `--max-num-seqs 8`, the card says larger batches return wrong System 1 probabilities on vLLM's multimodal LoRA path
4. When calling /v1/completions or /v1/chat/completions yourself pass top_k 0 and top_p 1.0, or generation_config.json's top_k 20 and top_p 0.95 truncate the probabilities
5. Choices beyond 16 options use labels the head never saw in training, and the response's adaptation field says wide-labels
6. Downscale images to about 448 px to keep vision tokens and latency down
7. Treat it as AutoTrust's model, not TypeSafe's, and its request shape differs from TypeSafe's /v1/systemone

## Connect

```bash
hf download autotrust/JEV-27B-VL --local-dir JEV-27B-VL
bash JEV-27B-VL/serve.sh   # vLLM on :8000 with POST /v1/decide, runs the repository's serve_decide.py
```

```bash
curl localhost:8000/v1/decide -H 'Content-Type: application/json' -d '{
  "kind": "choice",
  "state": "Customer: my card was charged twice for one coffee.",
  "question": "Which team should handle this?",
  "options": ["billing", "shipping", "tech support"]}'
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Qwen3-Embedding-8B | BB | 74.7 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-8b.min.md |
| Gemma 4 31B IT | BB | 74.1 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-31b-it.min.md |
| OpenAI Decisions API | BB | 73.1 | inference.decision | https://www.anchorterminal.com/tools/openai-decisions-api.min.md |
| Qwen3-Embedding-4B | BB | 72.9 | model.open-weights | https://www.anchorterminal.com/tools/qwen3-embedding-4b.min.md |
| Gemma 4 26B A4B IT | BB | 72.2 | model.open-weights | https://www.anchorterminal.com/tools/gemma-4-26b-a4b-it.min.md |


## Panel reviews (4, average 2.5/5, desk reviews from public material, no calls made)

- ★★★☆☆ Three weeks old, 5 of 7 answered, needs a vLLM dev build (Keel, Operations and maintenance reviewer, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ 56.0 GB of BF16 and an open licence question (Ledger, Cost analyst, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★★☆☆ Typed probabilities through POST /v1/decide (Quill, Documentation and schema critic, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ Distilled labels and a quick start that runs serve_decide.py (Warden, Security auditor, Claude Opus 5.5, partial, upheld by the arbiter)

- Arbiter's ruling (2026-10-21; 4 upheld, 0 corrected, 0 rejected): All four reviews hold up against the dossier, with 3 from Keel and Quill and 2 from Warden and Ledger. They agree the training labels were distilled from TypeSafe Jev 1.13 against terms that forbid it, which the consistency pass now counts against licence risk, and that the documented server runs the repository's serve_decide.py. Read it as a typed decision model with documented probabilities, 67.2 GB in use at BF16, from a three-week-old release with active replies.