Interfaze: A New Model Architecture Delivering High Accuracy at Scale

Interfaze presents a hybrid architecture combining DNN/CNN + transformer decoder, explaining that it is tailored for deterministic tasks such as OCR, object/GUI detection, web extraction/search, STT, and translation.

• Specs: context window 1M tokens, max output 32k tokens, input supports text/image/audio/file, with reasoning disabled by default.
• Own benchmarks: it published comparison results against Gemini-3-Flash, Claude-Sonnet-4.6, GPT-5.4-Mini, and Grok-4.3 on OCRBench V2, olmOCR, RefCOCO, VoxPopuli-Cleaned-AA, Spider 2.0-Lite, GPQA Diamond, MMMLU, MMMU-Pro, and SOB.
• Pricing: set at input $1.50 / 1M tokens, output $3.50 / 1M tokens.
• OCR: it explains that it returns pixel coordinates for both text and graphical objects together from complex PDFs and images, and even exposes raw OCR precontext and confidence.
• STT: it presented WER 2.4% on VoxPopuli-Cleaned-AA, and stated it processes 209 seconds of audio per second.
• Integration: it can be used by connecting an OpenAI Chat Completions-compatible API at https://api.interfaze.ai/v1.

SOB is explained as a benchmark that measures the model's ability to accurately fill in values in a JSON structure when the correct answer is already given in the context.