Fifteen days of work, about $1,200 in rented GPU time and a 4-billion-parameter model: ImaJev-4b has entered the leaderboards with first place on JevBench (67.37 versus Jev 1.13.0 at 63.29) and third place on DecisionBench, ahead of GPT-5.6 Luna and DeepSeek V4.1. The news is not the score itself, but what it says about the possibility of bringing complex business decisions to compact models that run locally.

The creator starts from concrete experience: as a process consultant for Fortune 500 companies, he saw decision nodes around refunds, returns and customer support remain assigned to people because code struggles to handle ambiguity. The arrival of Jev made the use case clear, but without image support. Hence the ImaJev bet: a single model for text and photos, with answers to closed questions and a probability for each option.

The path was not linear. An initial fine-tuning on about 500,000 short decisions made a 9-billion-parameter model worse: on JevBench hard the score dropped from 64.9 to 42.3, a sign of pattern matching rather than reasoning. The recovery came only after weeks of generating difficult questions with open-weight models and keeping only cases where two models agreed. This detail may be the most important: data quality reversed the outcome, not model scale.

Technically, ImaJev-4b uses LoRA and a small decision head on Qwen3.5-4B. It accepts text or a JSON record, up to two photos and closed questions. In a single forward pass it returns probabilities for each option plus an 'unknown' class. It runs on a Mac with MLX or on a single GPU, with weights distributed under Apache-2.0.

Here is where the analysis goes beyond the benchmarks. The first place on JevBench comes from a metric that weights accuracy, calibration, speed and cost equally. On accuracy alone, ImaJev-4b is third. Its real strength is different: when it assigns a 90% probability it tends to be right, and when uncertain it says 'can't tell' instead of guessing. In a business process this changes the economics of decision-making: calibrated abstention can send a case to human review, avoiding costly errors and reducing the load.

The stakes for on-premise deployment are structural. A specialized 4B model with LoRA, runnable on a single GPU or on a Mac, keeps transactional data and photos inside the company perimeter, with contained latency and cost. It does not require a fleet of cloud accelerators: it requires curated data, closed questions and a clear abstention policy. This shifts attention away from the 'bigger model' parameter toward training set quality and probability calibration.

To be sure, the limits are explicit: the model works on closed questions, accepts at most two images and does not promise open reasoning. Leaderboards can also reward composite metrics that do not equal production value. But for those managing repetitive decision processes, the signal is precise: compact open-weight models with targeted fine-tuning can own circumscribed decisions, leaving more general open tasks to larger systems. Those evaluating on-premise deployment of such flows can find analytical frameworks on AI-RADAR's /llm-onpremise to weigh the trade-offs, without shortcuts.

The final point remains the most interesting: after the failure of the first attempt, the quality jump did not come from more parameters, but from a dataset filtered for agreement between models. It is a reminder for anyone building decision pipelines: in fine-tuning, disagreement is noise, consensus is supervision.