What happened
AutoTrust AI Lab's Hugging Face page presents JEV-35B as a “System 1” decision model rather than a conversational assistant. It is based on Qwen3.5-35B-A3B, which the page describes as having 35 billion parameters and about 3 billion active per token. The model is intended to take a decision problem and return scores or probabilities for a set of options, with up to 256 options handled in one pass. Its quick-start section shows a vLLM-style POST /v1/decide interface, and the page also lists Transformers loading instructions using AutoProcessor and AutoModelForMultimodalLM.
The repository is public, ungated, and marked Apache-2.0. AutoTrust's page classifies it under text classification, while also tagging it as image-text-to-text, typed decisions, calibrated probabilities, mixture of experts, and vLLM. The listed BF16 safetensors total about 72 GB, with roughly 35.95 billion BF16 parameters. Open weights therefore do not make this a small local model: serving the BF16 version requires substantial accelerator memory and an appropriate runtime.
The card also points to a separate NVFP4 variant. AutoTrust says that version has a 25.6 GB download and needs 23 GiB of GPU memory for weights, compared with 72 GB download and 66.5 GiB for BF16. The publisher says the quantized model can run on one 32–48 GB GPU. Its reported Decision Index 0.3 score is 59.18 versus 59.73 for BF16, and its reported vision rebuild score is 71.23 versus 71.94. AutoTrust further reports that 96.6% of Decision Index answers were identical to BF16 and that latency and the serving interface were the same.
AutoTrust reports a 59.73 result on its Decision Index 0.3 public suite, explicitly describing it as the authors' scoring with the kit rather than a board entry; the repository marks that result unverified. The page lists several external-style evaluation records as unverified too, including 0.70 accuracy, 0.169 Brier score, and 0.0617 expected calibration error on Typed Decisions. It also reports a 51.3% ERQA result for one-pass embodied reasoning, despite saying that no robot or embodied data was used in decision training. JEV-35B has no thinking mode, according to the model card. Those figures are publisher-reported evaluations, not independent confirmation of reliability or transfer to production decisions.
The practical boundary is clear: this is a model for producing structured judgments among supplied alternatives, not a general-purpose chat endpoint. The release provides weights, a permissive licence, and serving directions. A team still has to define the candidate actions, construct the decision input, set acceptance thresholds, and decide what happens when the model expresses low confidence or is wrong.
What the archive adds
New Runtime's earlier coverage includes GEV-26B-Decide (12 October 2026) and AutoTrust-JEV-27B v1.1 (8 November 2026), both described as decision models that output probabilities across options. JEV-35B is a new release in that recurring thread, with a larger model and a stated 256-option capacity; it is not a new category of interface.
Trend signal
This supports the shift toward model answers becoming task interfaces: a bounded choice output can be more directly consumed by downstream software than free-form text. The fit is suggestive rather than proven. A probability vector can simplify integration, but the model card's unverified results do not show that its judgments are dependable in a particular operational setting.
Why it matters
A decision-only interface can help separate candidate generation from selection. One system can produce a fixed list of actions, documents, products, alerts, or routes; JEV-35B can then score that defined set; and application logic can send the result to a reviewer, a queue, or a bounded rule. This is potentially easier to instrument than extracting a decision from a prose answer. The output format makes it possible to log each candidate set, chosen option, confidence distribution, threshold, and final outcome.
That structure is useful only if the task is genuinely a finite choice problem. A poorly specified list can force a plausible answer into the wrong set of alternatives, while a high score can still be miscalibrated. The model is therefore not a replacement for task design or accountability. It is a possible selection component between generation and execution.
The operating costs and evidence caveats are material. BF16 weights require significant memory, while the lower-memory NVFP4 version carries publisher-reported score reductions. The headline Decision Index result is both publisher-scored and marked unverified. Teams should regard the release as a candidate for controlled evaluation rather than a ready-made authority. For consequential actions, compare its ranking and calibration with a simple baseline on representative cases, preserve abstention and human-review paths, and monitor results after deployment.
Limits and uncertainty
The repository marks the listed Decision Index result unverified and describes it as the authors' own scoring rather than a board entry. Its evaluation records are also labeled unverified. The reported ERQA result and comparisons with other systems are publisher claims; New Runtime has not independently reproduced those evaluations.
The stated probability outputs should not be presumed calibrated for every decision domain. Calibration can vary with the candidate set, prompt construction, language, distribution shift, and the cost of false positives relative to false negatives. The card's hardware figures describe listed weight formats, but actual throughput, latency, batching behavior, and serving cost will depend on the deployment configuration. The permissive licence governs the repository, not the suitability of the model for a regulated or high-stakes workflow.
Skill to build now
Build a small decision-evaluation harness this week. Select one recurring workflow with a fixed action set, such as triaging support tickets into four queues or ranking five retrieval candidates. Save the inputs, option lists, JEV-35B scores, chosen action, reviewer correction, and eventual outcome. Calculate top-choice accuracy and a calibration measure before connecting the output to automation. Add an abstain rule for narrow score margins, and compare it with a basic heuristic or an existing model. This turns a probability-shaped API into a testable operational component.
Reader takeaway
JEV-35B makes AutoTrust's decision-model approach available as Apache-2.0 open weights, with a stated capacity of up to 256 options in one pass. Its value is not that a score looks authoritative: it is that selection can become an explicit, measurable interface. Treat the probabilities as hypotheses to validate on your own decisions, while planning for the model's substantial hardware demands.