FinTech News & Updates

Stay informed with the latest breaking news, market updates, and industry developments

Plaid Decagon partnership brings bank verification into AI customer support
Sep 29, 2026 – FinTech News

Plaid and Decagon Bring Bank Verification Into AI Customer Support

Customer support agents are increasingly moving beyond answering questions. Plaid and Decagon are bringing bank-account connectivity directly into AI-powered customer…

Read More →
HSBCnio brings AI and APIs into corporate transaction banking workflows
Sep 29, 2026 – FinTech News

HSBC Launches HSBCnio to Embed AI and APIs Into Corporate Transaction Banking

HSBC is bringing AI, APIs, and transaction banking services directly into corporate workflows with the launch of HSBCnio, a new…

Read More →
SWIFT cross-border payments connecting Bizum, PayID and Pix through pay-by-alias technology
Sep 29, 2026 – FinTech News

SWIFT Connects Bizum, PayID and Pix to Simplify Cross-Border Payments

Sending money across borders could soon feel much closer to making a domestic instant payment. SWIFT is working with payment…

Read More →
Marble fraud AML platform raises €6.5M Series A to expand AI compliance technology
Sep 29, 2026 – FinTech News

Marble Raises €6.5M to Expand Open-Source Fraud and AML Platform

Paris-based fraud and anti-money laundering technology company Marble has raised €6.5 million in a Series A round led by Smartfin.…

Read More →
Apple Pay antitrust case expands to thousands of U.S. banks over payment fees
Sep 29, 2026 – FinTech News

Apple Pay Antitrust Case Expands to Thousands of U.S. Banks and Credit Unions

Thousands of U.S. banks and credit unions can now pursue claims against Apple together over fees charged on Apple Pay…

Read More →
NVIDIA Open Agent Safety Platform securing AI agents across software and hardware
Sep 29, 2026 – FinTech News

NVIDIA Launches Open Agent Safety Platform to Control AI Agents From Software to Silicon

NVIDIA has launched an Open Agent Safety Platform designed to give companies greater control over autonomous AI agents, from testing…

Read More →
Meta AI spending concerns as Zuckerberg's fortune falls after Meta shares slide
Sep 29, 2026 – FinTech News

Zuckerberg Loses $10.9 Billion as Meta Shares Slide on AI Spending Concerns

Mark Zuckerberg‘s estimated fortune fell by $10.9 billion on Monday as Meta Platforms shares declined. This extended a two-session selloff…

Read More →
# On-Device AI Challenges the Cloud Model With Smaller, Faster AI Systems The next phase of artificial intelligence may not depend entirely on larger models and bigger data centers. On-device AI is emerging as an alternative approach, bringing increasingly capable models directly to smartphones, edge computers, and other local hardware. Several factors are driving this shift. Privacy requirements, latency, connectivity limitations, and energy consumption are pushing developers to explore smaller models that can perform specific tasks without sending data to the cloud. UNESCO and University College London research has highlighted the potential energy savings here. A UNESCO report says that using smaller models tailored to specific tasks can cut energy consumption by up to 90%, while model compression can deliver additional savings on top of that. The development of compact vision-language models shows just how quickly this approach is progressing. ## Why smaller AI models matter Much of the AI industry's infrastructure strategy has focused on increasingly large models that require substantial computing power in centralized data centers. That approach remains important for complex workloads. Still, not every AI task needs a massive general-purpose model. A smartphone application that needs to read a document, identify objects in a photograph, or answer a question about a scene may not require the same infrastructure as a large cloud-based reasoning system. Smaller models can instead be optimized for specific tasks and hardware. This creates several potential advantages. Privacy is one of the clearest. When processing happens locally, sensitive images or other inputs don't necessarily need to travel to a remote server. Latency is another. An application that processes information locally doesn't have to wait for a round trip between a device and a cloud server. Connectivity also becomes less important, since local AI can keep operating when an internet connection is weak or unavailable. There's also an infrastructure argument here. If more AI workloads can run efficiently on existing consumer devices, some applications may need less centralized computing capacity overall. ## VisionPsy-Nano brings a 460 million-parameter model to devices Tether AI Research released VisionPsy-Nano in July 2026 as an open-weight family of vision-language models designed for on-device and edge deployment. The model contains approximately 460 million parameters, putting it in a significantly smaller class than many large multimodal AI systems. Tether says the model handles visual perception, document understanding, reasoning, and instruction following directly on consumer hardware. The model comes in two variants. VisionPsy-Nano-460M prioritizes quality, while VisionPsy-Nano-460M-Flash is optimized for latency and lower-resource deployment instead. According to Tether's published benchmark results, the standard VisionPsy-Nano model compared favorably with other models in the roughly 0.5-billion-parameter class on 16 of 17 public benchmarks. The testing compared it against models including Liquid AI's LFM2.5-VL-450M and Hugging Face's SmolVLM2-500M. These figures come from Tether's own benchmarks, so readers should view them in that context. The model documentation states that the evaluations used a common VLMEvalKit setup, while some evaluation changes remained under review at the time of publication. ## The performance challenge to larger models The most interesting part of compact AI isn't simply that models are getting smaller. It's that developers are trying to preserve useful performance while cutting computational requirements at the same time. Tether's published results offer one example. VisionPsy-Nano-460M scored ahead of several larger models on selected benchmarks. The company's documentation reports that the 460M model outperformed models such as FastVLM-0.5B, Qwen3.5-0.8B, and InternVL3.5-1B on ScienceQA, MM-IFEval, and POPE. Results still vary by task, though. VisionPsy-Nano doesn't outperform larger models on every benchmark. Tether's documentation identifies MMVet as a clear exception among the 17 evaluations it reports. That distinction matters, because model size remains only one part of AI performance. Architecture, training data, optimization techniques, and the specific task can all shape the final results. ## Faster responses could change mobile AI For many device-based applications, latency can matter just as much as raw model accuracy. A navigation assistant, augmented-reality application, or robotic system may need to interpret its surroundings continuously. Waiting for information to travel to a cloud server and back can create a meaningful delay in these settings. Tether designed VisionPsy-Nano-460M-Flash specifically around this problem. According to Tether, the Flash version retains about 99% of the full model's normalized score while delivering substantially faster first-token generation in its testing. On devices including the Pixel 9, Galaxy S23, and Galaxy S25 Ultra, Tether reports first-token generation roughly 19 to 23 times faster than the compared models. On an iPhone 15, the company reports gains of up to about 36 times. Again, these figures come from vendor-reported benchmarks rather than independent performance measurements. Still, the direction matters. If smaller models can deliver acceptable quality with much lower latency, developers gain another option between fully cloud-based AI and conventional software without AI capabilities at all. ## Privacy becomes a core advantage Local processing also changes where data actually travels. With a cloud AI application, an image, document, or question may need to leave the device and reach a remote server for processing. An on-device model can instead process that information locally, without that extra step. This matters particularly for sensitive documents, personal photographs, financial information, and other data users may not want to transmit unnecessarily. That said, this approach doesn't automatically make every AI application private or secure. Developers still need to think through the operating system, application permissions, model integrity, storage, and other security controls carefully. But local inference can reduce the need to send certain inputs to external infrastructure at all. That makes on-device AI particularly interesting for applications where privacy is a core product requirement, rather than an optional feature bolted on later. ## Energy efficiency adds another reason to go small The case for smaller AI also extends well beyond smartphones. UNESCO's research found that several changes in how AI models get built and used can substantially reduce energy consumption. The organization says smaller models tailored to specific tasks can cut energy use by up to 90%, while shorter prompts and responses and model compression can reduce consumption further still. That doesn't mean every small model will automatically use 90% less energy than every large model. The actual result depends on the model, workload, hardware, and deployment environment involved. The broader point is simply that AI efficiency can improve by matching the model to the task at hand. For companies building AI products at scale, that could influence infrastructure costs as well as environmental considerations down the line. ## From smartphones to robotics On-device vision models could eventually find applications well beyond mobile assistants. Robotics is one clear example. A robot operating in a physical environment needs to interpret objects, people, and spatial relationships continuously. Sending every visual input to a distant cloud service can introduce latency and connectivity dependencies that simply don't work in real-world settings. Smaller vision-language models could let more of that perception and reasoning happen directly on the robot or at the edge instead. The same principle applies to augmented reality and other systems that interact with the physical world in real time. These applications may accept some reduction in general-purpose reasoning in exchange for faster response times and lower computing requirements. That trade-off looks quite different from the traditional AI race, where larger models have often been linked with greater capability by default. ## The cloud is not disappearing The growth of on-device AI doesn't mean cloud AI infrastructure is becoming unnecessary. Large models remain useful for complex reasoning, large-scale data processing, training, and workloads that exceed what consumer hardware can handle. Instead, the industry may be moving toward a more distributed model. Some tasks can stay in the cloud. Others can run on smartphones, PCs, vehicles, industrial equipment, or robots directly. Applications can also combine local and cloud inference depending on how complex a given request turns out to be. That hybrid architecture could become increasingly important as AI moves from chat interfaces into everyday physical environments. ## The bigger AI shift may be about efficiency The most important development in on-device AI isn't simply that smartphones can run smaller models now. It's the growing recognition that AI capability doesn't always require maximum model size. UNESCO's research points toward smaller, task-specific systems as one route to lower energy consumption, while Tether's VisionPsy-Nano offers a current example of developers optimizing a compact multimodal model for local execution. For users, the benefits could include faster responses, greater privacy, and AI functionality even in situations where an internet connection is unavailable. For developers, the attraction runs broader still: lower latency, more control over data flows, and the possibility of deploying AI across hardware with limited computing resources. The cloud will remain central to AI for the foreseeable future. But as smaller models become more capable, the smartphone and other edge devices are increasingly becoming part of the AI infrastructure themselves, rather than just endpoints that talk to it. --- On-device AI running smaller and faster AI models on smartphones and edge devices
Sep 28, 2026 – FinTech News

On-Device AI Challenges the Cloud Model With Smaller, Faster AI Systems

The next phase of artificial intelligence may not depend entirely on larger models and bigger data centers. On-device AI is…

Read More →
NVIDIA expands its share buyback authorization by $150 billion to $235 billion
Sep 28, 2026 – FinTech News

NVIDIA Expands Share Buyback Authorization by $150 Billion, the Largest Increase in History

NVIDIA has expanded its share repurchase program by $150 billion. This move gives the AI chipmaker $235 billion in remaining…

Read More →
PB Fintech stock rout after IRDAI proposes insurance commission caps
Sep 28, 2026 – FinTech News

PB Fintech’s $3.3 Billion Market-Value Rout Drives Surge in Leveraged Positions

PB Fintech‘s sharp stock selloff has been followed by a surge in leveraged positions in the company. Traders increased margin-funded…

Read More →
1 … 26 27 28 29 30 … 375