Six Tiny AI Models That Run Entirely in Browser or on CPU, With Source Codes
Small models that run on CPU or in the browser make AI features like background removal, OCR, and face recognition practical for client-side apps without GPU costs or Python backends. This collection shows the current size and performance bar: several tasks fit in under 10MB.
A developer who has been publishing small, deployable AI models has collected six of them into one post with download links. Every model runs in a web page or locally on CPU, with model sizes ranging from 5.5MB to 582MB. These cover background removal (46MB), OCR (6MB), old photo restoration (Tencent's 582MB model), torso recognition (Google's 5.5MB model), face recognition (6.5MB), and speech-to-text via Alibaba's ASR.
The post is a straightforward download roundup aimed at readers who want the source code without GPU or Python setup. Web-based implementations are emphasized, and the author notes that the background removal tool now removes arbitrary subjects. A server-side background removal model (SAM3.1, 3.21GB) is teased for release later this week.
The collection points to a growing trend of small, self-hostable AI tools that fit in a browser tab or a lightweight local runtime. Baidu's OCR model is called out as surprisingly fast and accurate, and Google's torso model handles video without lag.
Bundling six separate tasks into one post signals that edge AI is moving from one-off demos to a toolbox of deployable utilities.
The emphasis on web pages and CPU-only deployment fits a user base that wants to avoid cloud costs and data privacy trade-offs.
Model sizes under 10MB for OCR, torso, and face recognition suggest that on-device inference for common vision tasks is now cheap enough for casual web pages.