AI/TLDR

Interfaze · 2026-05-11 · notable

Interfaze — Hybrid CNN/DNN + LLM Architecture Beats Gemini-3-Flash, Claude-Sonnet-4.6, GPT-5.4-Mini, Grok-4.3 Across 9 OCR, Vision, STT, and Structured-Output Benchmarks

Y Combinator-backed Interfaze (formerly JigsawStack) details a hybrid model that routes deterministic developer tasks to specialized CNN/DNN models behind an LLM front-end, claiming head-to-head wins over flagship Flash/Mini-tier models.

Interfaze architecture blog post social card with the title 'A new model architecture built for high accuracy at scale'

A hybrid model that bolts CNNs and small task-specific networks under an LLM front-end and routes deterministic developer jobs to whichever specialized model is best.

Key specs

Context window1M tokens
Ocrbench v2 score70.7%
Olm ocr score85.7%
Vox populi stt wer2.4%
Max output tokens32,000
Input price per million tokens$1.50
Output price per million tokens$3.50

What is it?

Interfaze is a new model architecture from JigsawStack (the company has rebranded to match the product name). Instead of a single monolithic transformer, the system fronts a suite of small specialized models — OCR networks, vision-language models, speech-to-text models, structured-output extractors — with a controller that routes each request to the right specialist. The team positions Interfaze for deterministic developer tasks (OCR, scraping, classification, STT, structured output) rather than open-ended chat.

How does it work?

The architecture has four stages: an ingress that classifies the request, a small-model and perception stack that runs the specialist (CNN/DNN for vision and OCR, dedicated STT models for audio), a context-construction layer that assembles structured prompts, and a thin action layer driven by an LLM-style controller. Output is constrained to schemas where applicable, with optional reasoning. The team reports the hybrid stack beats flagship Flash/Mini-tier models on nine benchmarks including OCRBench V2 (70.7% vs Gemini-3-Flash's 55.8%) and olmOCR (85.7% vs Claude-Sonnet's 73.9%).

Why does it matter?

Most LLM APIs treat OCR, structured extraction, and STT as second-class — they work, but accuracy degrades sharply at long inputs or rare formats. A vendor selling specialized small models behind a single API at $1.50/$3.50 per million tokens is a sharper price-performance offer than calling Gemini-3-Flash for the same job. The post is the first sustained pitch of the hybrid approach since the JigsawStack rebrand.

Who is it for?

Developers running pipelines on OCR, scraping, classification, structured output, or audio transcription

Try it

https://interfaze.ai/

Sources · 3 outlets

Tags

  • interfaze
  • jigsawstack
  • hybrid-architecture
  • ocr
  • structured-output
  • speech-to-text
  • function-calling
  • developer-tools
  • y-combinator

← All releases · Learn AI