On-tablet object recognition neural network hits 94% F1-score in 20 ms
YADRO has described how it taught its KVADRA_T tablet to detect a person, a document, text, QR codes and barcodes in the same frame at once. Instead of a standard multiclass setup, the team chose a multi-label approach, built a dataset of 193,000 images and reached an F1-score of 94%. Inference on the tablet's NPU takes about 20 ms, and the feature is expected to ship in the next kvadraOS release.
AI-processed from Habr AI; edited by Hamidun News
YADRO has developed a neural network that recognizes a person, a document, text, QR codes and barcodes simultaneously on the KVADRA_T tablet, reaching an F1-score of 94% and a processing time of 20 milliseconds directly on the device. The company explains that standard multiclass classification did not fit the task by its very nature: a single image can contain several objects of different types at once, so a multi-label model was needed.
Why doesn't standard classification work for recognizing multiple objects?
A single frame can contain a person, a passport, lines of text and scannable codes at the same time, while classic multiclass classification assumes that an object belongs to only one class. The task calls for a multi-label model, which recognizes all the objects at once.
How does the neural network process data in 20 ms on a tablet without the cloud?
The model is optimized to run directly on the device (on-device inference), which lets it process frames in 20 milliseconds without an internet connection. That is a requirement for practical use on mobile devices.
- Nov 1, 2026Local models lose to Opus and Codex: 0.21 on a MacBook versus 1.00 for Codex
- Nov 1, 2026MoneyGate safety layer blocked all 42 fraudulent requests to local AI models
- Nov 1, 2026ThinkJev Makes AI Think Only About Hard Questions: 28% Fewer Tokens
- Oct 31, 2026Open-source LLM gateway KTMS 1.0 plugs your own GPUs into the cloud and survives outages
Key Questions Answered
Why doesn't standard classification work for recognizing multiple objects?
A single frame can contain a person, a passport, lines of text and scannable codes at the same time, while classic multiclass classification assumes that an object belongs to only one class. The task calls for a multi-label model, which recognizes all the objects at once.
How does the neural network process data in 20 ms on a tablet without the cloud?
The model is optimized to run directly on the device (on-device inference), which lets it process frames in 20 milliseconds without an internet connection. That is a requirement for practical use on mobile devices.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.