# PyTorch 2.13 Adds FlexAttention for Apple Silicon: 12x Speedup on Sparse Attention

> PyTorch 2.13 brings FlexAttention to Apple Silicon, and the number attached is eye-catching: up to a 12x speedup on sparse attention patterns. Attention is the core operation inside modern AI models and also one of the most expensive, so making it dramatically faster on Apple's M-series chips matters for anyone who wants to run capable models on a laptop instead of renting a cloud GPU. Sparse attention is the trick being accelerated. Instead of having every token attend to every other token, which grows costly as inputs get longer, sparse patterns let a model skip the connections that do not matter. FlexAttention gives developers a flexible way to express those patterns and, now, to run them efficiently on consumer hardware. The quiet significance is on-device AI. Every gain that lets a strong model run locally chips away at the assumption that serious inference has to live in a data center, with implications for cost, privacy and who controls the compute. It is a plumbing update, not a headline model, but this is the kind of change that decides where AI actually runs.

_Section: [Daily AI Updates](https://www.wortins.com/daily-ai) · Source: AIapps · Published Thursday, July 16, 2026_

## Wortins' read

PyTorch 2.13 brings FlexAttention to Apple Silicon, and the number attached is eye-catching: up to a 12x speedup on sparse attention patterns. Attention is the core operation inside modern AI models and also one of the most expensive, so making it dramatically faster on Apple's M-series chips matters for anyone who wants to run capable models on a laptop instead of renting a cloud GPU. Sparse attention is the trick being accelerated. Instead of having every token attend to every other token, which grows costly as inputs get longer, sparse patterns let a model skip the connections that do not matter. FlexAttention gives developers a flexible way to express those patterns and, now, to run them efficiently on consumer hardware. The quiet significance is on-device AI. Every gain that lets a strong model run locally chips away at the assumption that serious inference has to live in a data center, with implications for cost, privacy and who controls the compute. It is a plumbing update, not a headline model, but this is the kind of change that decides where AI actually runs.

## Source

[Read the full story at AIapps](https://www.aiapps.com/blog/july-ai-mega-update-major-breakthroughs-launches/)

## Related coverage

- [South Korea announces 1,350 trillion won ($880B) decade-long AI infrastructure plan](https://www.wortins.com/story/south-korea-announces-1-350-trillion-won-880b-decade-long-ai-dd479678) — [Tech Startups](https://techstartups.com/2026/07/16/top-tech-news-today-july-15-2026/)
- [xAI open-sources Grok Build coding agent](https://www.wortins.com/story/xai-open-sources-grok-build-coding-agent-7997c3dd) — [Future Tools](https://futuretools.io/news)
- [Helsing Raises $1.8B Series E at $18B Valuation in Europe's Largest Defense-Tech Round](https://www.wortins.com/story/helsing-raises-1-8b-series-e-at-18b-valuation-in-europe-s-la-32ef801c) — [cnbc.com](https://www.cnbc.com/2026/07/13/helsing-fund-raise-defense-18-billion.html)
- [Andrej Karpathy and Tom Blomfield Join Anthropic's Team](https://www.wortins.com/story/andrej-karpathy-and-tom-blomfield-join-anthropic-s-team-0d91a963) — [unrot.co](https://unrot.co/blogs/today-top-10-ai-news-july-15-2026/)
- [PrismML releases Bonsai 27B, 27B multimodal model runs on iPhones](https://www.wortins.com/story/prismml-releases-bonsai-27b-27b-multimodal-model-runs-on-iph-0f594386) — [MarkTechPost](https://marktechpost.com/2026/07/16/prismml-releases-bonsai-27b-a-27b-compact-multimodal-model-available-in-binary-and-ternary-variants/)
- [Apple Intelligence approved for China, uses Alibaba Qwen models](https://www.wortins.com/story/apple-intelligence-approved-for-china-uses-alibaba-qwen-mode-808e0ae5) — [Build Fast with AI](https://www.buildfastwithai.com/blogs/ai-news-today-july-16-2026)

---

_Curated and written by [Wortins](https://www.wortins.com) — The daily AI briefing. Every story links to its original source; the "Wortins read" on each is our own original analysis. [About Wortins & our editorial approach](https://www.wortins.com/about)._