Safari 27.3 Brings Adaptive Dark Mode

// 1h agoPRODUCT UPDATE

Safari 27.3 Brings Adaptive Dark Mode

Safari 27.3 beta adds Adaptive Dark Mode, automatically transforming bright websites into readable dark versions while preserving images and respecting sites with native dark themes. The feature is built into WebKit and available for testing now. [WebKit announcement](https://webkit.org/blog/18541/introducing-adaptive-dark-mode/)

// ANALYSIS

Apple is effectively building Dark Reader-style functionality into Safari, turning a popular extension category into a browser-level capability.

–Color transformations preserve hue while targeting WCAG’s 4.5:1 contrast ratio.
–SVG logos, gradients, shadows, buttons, and text receive context-aware adjustments instead of a blunt full-page inversion.
–Developers can signal native dark-mode support with `color-scheme: light dark`, while `light only` is intended as an opt-out.
–Websites cannot detect the transformation through JavaScript or canvas, reducing fingerprinting and analytics complications.
–The feature should reduce demand for dark-mode extensions, but complex legacy layouts may still require developer testing and cleanup.

// TAGS

safaribrowser-extensiondevtoolcssaccessibilitydark-mode

DISCOVERED

1h ago

2026-11-04

PUBLISHED

4h ago

2026-11-03

RELEVANCE

5/ 10

AUTHOR

paddingtone

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

MODEL45m ago

OrcaRouter LFM3-1.7B-abliterated Slashes Refusals

OrcaRouter’s refusal-reduced derivative of Liquid AI’s 1.73B LFM3 model reports average harmful-prompt refusals falling from 87.3% to 2.8%, while retaining its 256K context and local-agent focus. The underlying LFM3 model is designed for roughly 1GB 4-bit deployment, but this repository currently ships only larger BF16 safetensors.

BENCHMARK49m ago

Qwen3.8-27B Uncensored Tops 110 t/s Locally

A community benchmark reports Qwen3.8-27B-Uncensored reaching 110 tokens/s decode and up to 1,000 tokens/s prefill on one Radeon RX 7900 XTX during agentic workloads. The result highlights how optimized local inference makes a 27B open-weight model practical on consumer hardware.

NEWS1h ago

Claude Opus 5.5 Raises Recursive Self-Improvement Stakes

Claude Opus 5.5 improves long-horizon coding while costing 40% less than Opus 5, but Anthropic warns future models may soon meet its threshold for AI-accelerated development.