Google has launched LiteRT.js, a new JavaScript tool that runs AI models right in your browser. Released on July 15, 2026, LiteRT.js works with .tflite models using WebAssembly, WebGPU, and an experimental Web Neural Network API. This means AI can run faster and smoother without leaving the browser.
LiteRT.js boosts speed by up to three times compared to other web runtimes. On a 2024 MacBook Pro with M4 Apple Silicon, using the GPU or neural processing unit (NPU) speeds up AI tasks by 5 to 60 times over just the CPU. This makes AI-powered apps more responsive and efficient.
There are some quirks. LiteRT.js doesn’t support splitting AI tasks between CPU and GPU. Also, it doesn’t clean up memory automatically. Developers must delete tensors explicitly to avoid memory leaks. Still, the speed gains make it a powerful tool for web-based AI.
Gemini AI Comes to Chrome in the UK
Just a day before LiteRT.js, Google rolled out Gemini in Chrome to UK users on July 14, 2026. Gemini is an AI assistant that can access your open browser tabs to answer questions and provide summaries. It shares tab content and URLs by default, but you can change settings to limit this.
Gemini can summarize web pages, compare info across multiple tabs, and even work with Google apps like Gmail, Maps, Calendar, and YouTube. It also includes Nano Banana, Google’s own AI image generator. This makes browsing smarter and more interactive. Google says, “Tabs are shared for more relevant answers. Your current tab is shared in new chats.”
AI That Remembers More and Runs Locally
Tether’s AI Research Group developed technology to help laptops remember more data without relying on the cloud. This helps with privacy and offline use. Meanwhile, PrismML announced Bonsai 27B, a 27-billion parameter AI model that can run on phones, launched on July 14, 2026.
Bonsai 27B comes in two versions: a 1-bit and a 1.58-bit ternary variant. Both are based on Qwen3.6 27B and excel at reasoning, coding, and agentic workflows. The 1-bit version runs at 11 tokens per second on an iPhone 17 Pro, using just 3.9GB of memory. Both variants were trained on Google v5 TPUs.
PrismML made Bonsai 27B freely available under the Apache 2.0 license. This opens up powerful AI tools to developers and hobbyists who want advanced models working on phones without big servers.
FPGA AI Acceleration Made Easier
On the hardware side, Microchip Technology released VectorBlox 3.0 SDK on July 14, 2026. This free software development kit helps deploy convolutional neural networks (CNNs) on PolarFire FPGA and SoC platforms. It supports sparse neural networks to cut down power use and computing demands.
VectorBlox 3.0 lets multiple AI workloads run on one device. It targets power-constrained environments, making it ideal for embedded systems and edge AI. Federico Fontana, Head of Hardware Engineering at AIKO, said, “The combination of PolarFire SoC and VectorBlox creates a powerful synergy for deploying AI-powered autonomy solutions directly in orbit.”
Shakeel Peera, Microchip’s VP and GM, highlights how this SDK helps bring AI acceleration to new hardware types. This marks a step forward in making AI faster and more efficient outside traditional CPUs and GPUs.
These developments show a clear trend: AI is moving closer to users, whether in browsers, phones, or embedded devices. Faster, smarter, and more local AI is shaping the future of everyday tech.
Based on
- Google Releases LiteRT.js: A JavaScript Binding of LiteRT That Runs .tflite Models in Browsers via WebGPU — marktechpost.com
- Google Brings Gemini In Chrome To UK Users — engadget.com
- How developers can free AI from the data centres | WIRED — wired.com
- PrismML Announces 1-bit Bonsai 27B – The First 27B Model to Run on a Phone | Markets Insider — markets.businessinsider.com
- Microchip Advances Neural Network Implementation with VectorBlox™ 3.0 Accelerator SDK | Markets Insider — markets.businessinsider.com
