Interfaze new model architecture: 9 benchmarks surpass Gemini/Claude/GPT, focusing on high-precision tasks

Core conclusion

Interfaze this week released a new hybrid model architecture that combines the high accuracy of specialized DNN/CNN models with the flexibility of general-purpose Transformers. In 9 benchmarks covering OCR, visual understanding, speech transcription and structured output, Interfaze comprehensively crushed Gemini-3-Flash, Claude-Sonnet-4.6, GPT-5.4-Mini and Grok-4.3 at the same price.

Key Points

• Event: Interfaze releases hybrid architecture model to support OCR, visual understanding, speech-to-text (STT), structured output and knowledge retrieval
• Performance: Leads all peer models in 9 head-to-head benchmarks, including OCRBench V2, olmOCR, RefCOCO, VoxPopuli, GPQA Diamond, and more
• Pricing: Same price range as Gemini-3-Flash / GPT-5.4-Mini
• API compatible: natively supports Chat Completions API and can be directly connected with OpenAI SDK

Background: Why the DNN/CNN route deserves renewed attention

The Transformer architecture has kicked off the AI wave of the past two years, with almost all developer tools built on top of a general-purpose LLM. But the Interfaze team discovered a key problem: For deterministic tasks (OCR, table extraction, GUI detection, etc.), the efficiency of the general Transformer is much lower than the traditional DNN/CNN architecture.

CNNs/DNNs, from LeNet-5 to ResNet to CRNN-CTC, have always been the gold standard for OCR and vision tasks. They can be up to 100 times more accurate than general-purpose Transformers for specific tasks, and can also output metadata such as bounding boxes and confidence scores.

The innovation of Interfaze is that it does not give up the flexibility of Transformer, but embeds the dedicated modules of DNN/CNN into Omni-Transformer, activating specific task modules on demand.

Key Impact Comparison

Detailed explanation of 9 benchmark tests

Interfaze is compared comprehensively with similar models in 9 cross-modal benchmarks:

Of particular concern is SOB (Structured Output Benchmark) - Interfaze's self-developed evaluation method: put the correct answer into the model context and test the filling accuracy of the JSON output. This goes straight to the core pain point that many developers have with LLM: The model can output JSON according to the schema, but the filled values ​​are often wrong.

Detailed explanation of core competencies

OCR + visual positioning: two tasks completed with one request

The traditional approach is the concatenation of OCR service + target detection model + post-processing script. Interfaze can output the position coordinates of full-page text and image elements simultaneously in one request.

Speech to text: 1.5 times faster than Deepgram

In the VoxPopuli-Cleaned-AA test, Interfaze ranked second in word error rate (WER), processing 209 seconds of audio per second—about 1.5 times faster than Deepgram Nova-3 and more than 11 times faster than Gemini-3-Flash.

Deterministic task modular activation

Interfaze allows controlling activated model modules via system prompts, without having to load full weights to perform a single task - meaning lower latency and less computational overhead.

Actual usage scenario

1. Complex PDF Extraction: Replacing redundant document processing pipelines, OCR + structured extraction can be completed in a single request
2. GUI automated testing: Use target detection capabilities to locate UI elements and generate interactive scripts
3. Multi-modal content review: Detect text and visual content in images simultaneously
4. Real-time voice transcription: Supports audio for more than 1 hour, with timestamp output in chunks
5. Enhanced Search Agent: Built-in web index, supports structured answers combined with real-time search

Tool entry

The following AI tools and services are covered in this article: OpenAI, Claude, Gemini, Deepgram, GPT-5.4-Mini, Claude-Sonnet-4.6, Grok-4.3.

Internal link guidance

• Want to save money by running local models yourself? Watch the tutorial: How to run local AI models on M4 Mac with LM Studio: A complete 30-minute tutorial
• If you made money with Claude Code + AI SaaS, watch this: He used Claude Code + AWS to build AI SaaS, and his monthly income was $12,000 for 3 months

Reference sources

• Interfaze
• Hacker News