#on-device-ai

#on-device-ai

The key addition here is the LLM Booster and support for Google's Gemini Nano V3, which is designed to run on your phone. This new version of the chip still uses the NPU 880. The rest is the same. The CPU has its top Cortex-A725 core running at up to 3.4GHz. This core is equipped with 1MB of L2 cache. The second cluster is three A725s at up to 3.2GHz, each with 512KB of L2 cache. Finally, there are four A725s running at up to 2.2GHz with 256KB of L2 cache.

Gemini Intelligence can automate multi-step tasks and execute them entirely in the background on its own, including things like sourcing and transforming information and interacting with various apps and websites autonomously on your behalf.

On Wednesday, the Paris-based AI lab released two new speech-to-text models: Voxtral Mini Transcribe V2 and Voxtral Realtime. The former is built to transcribe audio files in large batches and the latter for nearly real-time transcription, within 200 milliseconds; both can translate between 13 languages. Voxtral Realtime is freely available under an open source license.

As demand grows for privacy-first enterprise AI that can run without sending sensitive data to the cloud, SpotDraft has raised $8 million from Qualcomm Ventures in a strategic Series B extension to scale its on-device contract review tech for regulated legal workflows. The extension values SpotDraft at around $380 million, the startup told TechCrunch, nearly double its $190 million post-money valuation following its $56 million Series B in February of last year.

The quest for edge AI seems central to Apple's future approach, and to support it the company will consider the acquisition of smaller AI firms who can deliver optimized, compressed AI models. The company also intends to work with third-party models to shrink and adapt them to work more fully on Apple's hardware. Doing so is important, as the more intelligence Apple can put at the edge, the more it can reduce demand on hosted cloud-based AI, which will reduce infrastructure costs.

The company, which is based in San Francisco and has an office in Pune, India, is targeting up to $35 million this year as it builds a royalty-driven on-device AI business. That growth has buoyed the company, which now has post-money valuation of between $270 million and $300 million, up from around $100 million in its 2022 Series B, Kheterpal said.

Apple cares about privacy. That is why the AI features will run either on-device or using Apple's Private Cloud Compute, not on Google's servers. It will be possible to use third-party cloud services for complex problems, if required. But the use of Apple's own servers means Google won't have direct access to user data.

Merriam-Webster just named "slop" its word of the year, defining it as "digital content of low quality that is produced usually in quantity by means of artificial intelligence." The choice is blunt, almost mocking, and it captures something that has been building for months: a collective exhaustion with AI hype that promises intelligence but delivers mediocrity. Over the past three months, that exhaustion has started bleeding into Wall Street.

Tiiny AI, a US-based deep-tech startup, has unveiled the Pocket Lab, officially verified as the "world's smallest personal AI supercomputer." This palm-sized device, no larger than a typical power bank, is capable of running large language models (LLMs) with up to 120 billion parameters entirely on-device, without relying on cloud servers or external GPUs. Designer: Tiiny AI At its core, the Pocket Lab aims to make advanced artificial intelligence both personal and private.

There's no way to recharge the ring. Migicovsky says he didn't want yet another gadget to charge every day, so instead, the Pebble Index has non-rechargeable silver oxide hearing aid batteries designed to last 2 years with average use. Once the device's battery is nearly dead, users will receive a notification in the app, and the idea is you'll buy a new Pebble Index-an idea that's easier to get behind knowing the ring costs just $75.

"Unlike traditional chat models that generate text-based responses, Computer Use Agent (CUA) models like Fara-7B leverage computer interfaces, such as a mouse and keyboard, to complete tasks on behalf of users," Microsoft said in a blog post. "With only 7 billion parameters, Fara-7B achieves state-of-the-art performance within its size class and is competitive with larger, more resource-intensive agentic systems that depend on prompting multiple large models."

Adding a large language model to Shortcuts means it's easier to build automations that can simplify your life. Here's how: How This Works Head to Apple Shortcuts, create a new shortcut, and you'll see "Apple Intelligence" as one of the listed applications that's supported. There are a few Actions related to text, allowing you to do things like proofread, summarize, and make a list from text. You also get the ability to create an image, if you want.

Frauds involving digital transactions accounted for more than half of all reported bank fraud in 2024 - 13,516 cases resulting in losses of ₹5.2 billion (about $58.61 million), according to the Reserve Bank of India (RBI). Online scams caused an estimated ₹70 billion (roughly $789 million) in losses in the first five months of 2025, the Ministry of Home Affairs said.

The laptop you use tomorrow will be like a Mac: Beautifully-designed, easy-to-use, highly secure, and packed with enough power to run artificial intelligence (AI) on device. It will possess advanced memory handling to optimize the use of that precious component, rather than squandering cash, heat sinks, and internal real estate on RAM that doesn't usually get used. Designed to optimize the OS it runs, it will integrate with your mobile devices, have a built-in tracker in case of loss,

The feature, virtually identical to Apple's Private Cloud Compute, comes as companies reconcile users' demands for privacy with the growing computational needs of the latest AI applications. Many Google products run AI features like translation, audio summaries, and chatbot assistants, on-device, meaning data doesn't leave your phone, Chromebook, or whatever it is you're using. This isn't sustainable, Google says, as advancing AI tools need more reasoning and computational power than devices can supply.

The new API accepts either text input or a combination of image and text input and produces a text response. According to Google, it enables the use of natural language prompts across a variety of mobile use cases, including short translation, image and text classification, content generation, and more: Prompt API moves beyond pre-built functionality to support custom, app-specific GenAI use cases, allowing you to create unique features with complex data transformation. Prompt API uses Gemini Nano on-device to process data locally, enabling offline capability and improved user privacy.

At its core, EmbeddingGemma serves as a text embedding model. It translates text, such as notes, emails, or documents, into specialized numerical codes called vectors. These vectors represent the meaning of the text in a high-dimensional space, allowing devices to grasp context rather than just matching keywords. This fundamental capability enables much more intelligent and helpful search, organization, and other AI functionalities, powering generative AI experiences directly on user hardware.

The browser would have more value if it included an on-device AI model that could run without requiring access to the internet, O'Donnell said. "This provides a channel through which they can get hundreds of millions of people to download their model," he said. In that scenario, the browser could access heavyweight AI models in the cloud to handle more demanding tasks.

At the core of the Lumex CSS platform is the Armv9.3 C1 CPU cluster, featuring Scalable Matrix Extension 2 (SME2) units. These units provide up to five times faster AI processing and three times improved efficiency compared with previous generations. The CPU cluster includes configurations such as C1-Ultra and C1-Pro cores, tailored to meet the demands of flagship devices. Micro-architectural improvements across cores contribute to an average 30% performance uplift and a 12% reduction in power consumption for daily mobile workloads.

Intel's upcoming "Panther Lake" processor for AI PCs could be a key building block in the transition of Windows 11 into an AI agent-driven OS, analysts said this week. Panther Lake will be available in PCs starting early in 2026. The chip meets the qualifications set by Microsoft for Windows Copilot+ PCs and provides the performance needed to run a new generation of AI applications, analysts said.

Starting with the Echo Dot Max, it features a dual-driver design consisting of a woofer and a tweeter. The new design integrated the speaker directly into the device housing instead of a separate module, which doubles the internal air space for fuller sound. The device can also tune the sound to your surrounding and Amazon claims it can produce 3x the bass of the 5th gen Echo Dot.

During the Snapdragon Summit on Maui, Cristiano Amon, CEO of Qualcomm, gave a glimpse into where the (mobile) ecosystem they provide with chips is heading. Qualcomm envisions a future in which AI moves from the cloud to your devices, taking care of everything for you in every possible way. Qualcomm invited us to attend the Snapdragon Summit, where two new chips were presented: a new smartphone and a new compute chip. The latter is primarily intended for laptops and mini PCs.

The way this works is silicon partners build and maintain execution providers that Windows ML distributes, manages, and registers to run AI workloads performantly on-device, serving as a hardware abstraction layer for developers and a way to get optimal performance for each specific silicon," Microsoft says in the announcement. In simple terms, the platform enables AI-infused apps to tap into PC hardware that's best suited for their specific workload, such as GPUs for power-intensive tasks, NPUs for power efficiency, and CPUs for flexibility.