AutoTrust Decider
AutoTrust's open decision models on Qwen and Gemma, 9B to 35B: give a state and typed questions, get a probability per option. Apache 2.0.
In short. AutoTrust Decider is a family of models that answer questions about a situation with a probability for each possible answer instead of writing text. You can send a support ticket and ask which team should handle it, and you get back a number for each team. The original Qwen or Gemma model stays untouched inside, so the same server can also write ordinary text.
Summary
AutoTrust AI Lab's first decision model on Hugging Face, Decider Mini 9B, dates from 23 September 2026, when its repository was created. The lab has since added Decider Pro 27B, a vision version of it, Decider Pro 35B and GEV-26B-Decide, all under Apache 2.0, and the model cards still use the older names JEV and GEV. They answer typed questions about a state, which is text, JSON or, in the vision versions, images. A question can ask for yes or no, one of 2 to 256 options or a 0 to 5 rating, and each answer is a calibrated probability for every option from one forward pass, with no generated text. The backbones are Qwen3.5-9B, Qwen3.8-27B, Qwen3.5-35B-A3B and Gemma 4 26B-A4B. AutoTrust leaves the original model untouched as what it calls System 2, for ordinary text generation and reasoning, and adds a small System 1 adapter and decision head, 108.9 million trained parameters in the 27B, so one vLLM server can do both. It says HumanEval stays at 78.0% for the 27B, identical to Qwen3.8-27B.
The numbers are AutoTrust's own. Its cards give an average of 84.07% over six public decision benchmarks for the 27B, against 83.85% for Typesafe's hosted Jev 1.13 in its own runs, and say the vision version ranked first of 20 on the community Jev Decision Index vision board, with a full score of 69.82 in a snapshot from 7 October 2026. A text-only v1.1 reports a public index of 67.31 on Decision Index 0.3, self-reported and pending verification by the index maintainers, and the index editions and settings differ between cards, so compare scores only within one card. The lineage matters. The System 1 labels come from SargeDev/jev-distill-corpus-v3, a dataset whose card says about 498,000 of its rows were distilled from the output distributions of Typesafe's closed Jev 1.13 through OpenRouter. AutoTrust reports a mean KL divergence of about 0.017 from that teacher for the 27B and says it is not affiliated with Typesafe AI, so check the terms that apply to your use.
The checkpoints are bf16 repositories of 17.9 GB for the 9B, 53.8 GB for the 27B, 55.6 GB for the vision 27B and 71.9 GB for the 35B, with NVFP4 builds of 25.6 GB and 18.8 GB for the 35B and for GEV-26B-Decide. The serve scripts run vLLM and add a POST /v1/decide route, the vision 27B's needs a vLLM development build, and the scripts for the vision 27B, the 35B and GEV ask for a GPU with 80 GB or more, or 32 GB or more for the 35B NVFP4. Prompts can reach 256K tokens for the vision 27B and GEV, and the card's single-fact test passed 20 of 20 at every length. AutoTrust measured a median of about 90 ms per decision for the 9B and 137 ms for the 27B on one B200, against 238 to 301 ms measured independently for the hosted Jev API. In the 30 days to 11 November 2026, Hugging Face counted about 1.65 million downloads of Decider Pro 35B, 1.55 million of the vision 27B and 675,000 of the 9B.
Good for
- Routing and triage steps in agents that need a fast, calibrated probability for each allowed option
- Intent classification, moderation and relevance ranking, with up to 256 options per question
- Judging an answer, an agent step or an image as a calibrated yes or no
- Teams that want typed yes or no, choice and 0 to 5 score questions next to ordinary text generation on one vLLM server
Watch out for
- The System 1 labels were distilled from Typesafe's closed, hosted Jev 1.13 through OpenRouter, so check the terms that apply to you even though the weights are Apache 2.0
- The scores are AutoTrust's own runs or self-reported results on a community index, and the board and its editions change over time
- Only bf16 weights are published for the 9B and the 27B models, so our 4-bit memory estimate is optimistic for them, and the larger models ask for an 80 GB GPU unless you use NVFP4
- It needs the repository's own server script, a vLLM development build for the vision 27B, and a cap of 8 concurrent sequences for that version
- The decision head was trained on text only, so image decisions are zero-shot and their calibration is not systematically measured
- Little is public about the organization, and the Hugging Face repositories and the model cards use different names
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Estimates for the Decider Mini 9B (JEV-9B) variant. A mixture-of-experts model must hold all of its experts in memory even though only some are used for each token, so total parameters decide the memory and active parameters decide the speed.
| Variants | Parameters | BF16 | FP8 / INT8 | 4-bit (Q4) | Fits on |
|---|---|---|---|---|---|
| Decider Mini 9B (JEV-9B) BF16 | 8.95B | 21 GB | 10 GB | 6.18 GB | Laptop or phone |
| GEV-26B-Decide (Gemma 4) BF16 | 25.8B 4B active | 59 GB | 30 GB | 80 GB | Workstation |
| GEV-26B-Decide NVFP4 NVFP4 | 25.8B 4B active | 59 GB | 30 GB | 21 GB | Gaming GPU |
| Decider Pro 27B (JEV-27B) BF16 | 26.9B | 62 GB | 31 GB | 19 GB | Gaming GPU |
| Decider Pro Vision 27B v1.1 (text decisions) BF16 | 27.4B | 63 GB | 31 GB | 19 GB | Gaming GPU |
| Decider Pro Vision 27B (JEV-27B-VL) BF16 | 27.8B | 64 GB | 32 GB | 80 GB | Workstation |
| Decider Pro 35B (JEV-35B) BF16 | 36B 3B active | 83 GB | 41 GB | 80 GB | Workstation |
| Decider Pro 35B NVFP4 NVFP4 | 36B 3B active | 83 GB | 41 GB | 32 GB | Gaming GPU |
Estimates: weights plus about 15% overhead. Real usage grows with context length and depends on the software. The developer states a minimum of 32 GB of GPU memory.
The license in plain words
One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.
See also
Frequently asked questions
- What is AutoTrust Decider?
- AutoTrust Decider is a family of models that answer questions about a situation with a probability for each possible answer instead of writing text. You can send a support ticket and ask which team should handle it, and you get back a number for each team. The original Qwen or Gemma model stays untouched inside, so the same server can also write ordinary text.
- Can I use AutoTrust Decider commercially?
- Yes. AutoTrust Decider is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
- Is AutoTrust Decider open source?
- AutoTrust Decider is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
- Can I run AutoTrust Decider on my own computer?
- Yes. The smallest variant, Decider Mini 9B (JEV-9B), needs about 6.18 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was AutoTrust Decider released?
- AutoTrust Decider was released by AutoTrust AI Lab on September 23, 2026.
- How long a context window does AutoTrust Decider have?
- AutoTrust Decider supports a context window of up to 256K tokens.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- AutoTrust Decider Pro Vision 27B model card (Hugging Face) huggingface.co
- AutoTrust Decider Pro 35B model card (Hugging Face) huggingface.co
- AutoTrust Decider Pro 27B model card (Hugging Face) huggingface.co
- AutoTrust Decider Mini 9B model card (Hugging Face) huggingface.co
- GEV-26B-Decide model card (Hugging Face) huggingface.co
- AutoTrust Decider Pro Vision 27B v1.1 model card (Hugging Face) huggingface.co
- Jev distill corpus v3 dataset card (Hugging Face) huggingface.co
- Jev Decision Index (Hugging Face Space) huggingface.co
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Clef Cloudflare | 9.41B – 27.4B | 64K | Apache 2.0 (Permissive) | 6.49 GB Laptop or phone | |
| pplx-decider Perplexity | 4.54B – 26.1B | 256K | Apache 2.0 (Permissive) | 3.13 GB Laptop or phone | |
| Liquid d1 Liquid AI | 587M – 3.12B | 128K | LFM Open 1.0 (Conditional) | 0.405 GB Laptop or phone | |
| Qwen3.8-27B Alibaba Qwen | 27.8B | 256K | Apache 2.0 (Permissive) | 19 GB Gaming GPU |