Interfaze: Hybrid DNN/Transformer Architecture Beats GPT-5.4-Mini and Claude Sonnet 4.6 on OCR Benchmarks — 146 Points on HN

Interfaze / Hacker News·medium signal

Interfaze published a new hybrid model architecture combining deep neural networks and CNNs with transformer decoders, targeting deterministic high-accuracy tasks over general-purpose reasoning, with a 1M-token context window and multimodal input support (text, images, audio). On OCRBench V2, Interfaze scored 70.7% vs. Gemini-3-Flash's 55.8%, and achieved 2.4% word error rate on VoxPopuli speech, at $1.50/$3.50 per million input/output tokens. The 146-point HN discussion examines whether reviving CNN/DNN approaches for structured tasks represents a viable alternative to pure transformer scaling.