Jev-style decisions for IMAGES: Fovea-4B, open 4B, one forward pass (huggingface.co)

Fovea-4B, a new AI model developed to enhance image understanding, can now answer typed questions about images in a single forward pass, providing probabilities for each answer. This model efficiently processes various types of images—such as documents, receipts, and charts—using TypeSafe's System One API. Its architecture is based on Qwen3.5-4B, featuring a hybrid attention mechanism combined with a Gated DeltaNet. Notable technical elements include using bfloat16 for weights, a decision readout in float32, and calibrated temperature settings tailored to different question types, enhancing the model's accuracy. This advancement is significant for the AI/ML community as Fovea-4B outperforms several leading models in the ImageJevBench v0.3 leaderboard, achieving an accuracy of 80.6% across 2,360 held-out multiple-choice questions. It shows notable improvements over Imajev-4B and H2O-Lightning-4B, making it a competitive choice for tasks requiring image interpretation. Moreover, the model’s ability to handle images up to 2 million pixels, combined with its efficient single-pass decision-making, presents a promising development for applications in document analysis, spatial reasoning, and multi-image assessments, pushing the boundaries of AI’s capability to integrate visual data and provide actionable insights.