Meta has spent much of this year trying to calm concerns about facial recognition coming to its smart glasses. But the company doesn’t look ready to give up on the idea, and a newly public patent application shows just how seriously Meta has continued exploring the technology.

First reported by Patentlyze, the application was filed by Meta in February and describes using AI-powered facial recognition alongside smart glasses to identify people and use information associated with them when creating media.

A patent filing does not guarantee any of these features will reach consumers. But it arrives after months of reports showing that Meta has repeatedly explored facial recognition for its wearable devices.

This is starting to look like more than an experiment

Back in February, reports revealed an internal feature called “Name Tag” that could allow Meta’s smart glasses to identify people in real time.

Things got considerably stranger in June. Researchers found dormant face-recognition components inside the Meta AI app, including systems capable of detecting faces and converting them into biometric data. Meta later removed the code and maintained that no facial-recognition feature had launched.

Around the same time, we also learned that Meta had licensed facial-recognition technology from Rank One Computing, a company whose customers include U.S. military and law enforcement agencies. The software was also found sitting dormant inside the Meta AI app.

At this point, there is no doubt that Meta has explored facial recognition for its glasses. What remains unclear is whether any of that work will eventually make its way into a product people can buy.

The privacy problem hasn’t changed

The patent has already drawn criticism. In a statement cited by CNET, the Center for Democracy and Technology warned that the technology could be abused in ways that put vulnerable people, including stalking and domestic violence victims, at risk.

It is the same concern that has followed Name Tag from the beginning. Smart glasses are very different from facial recognition tucked away inside a photo library. They are cameras worn on someone’s face while they move through public spaces.

Meta’s glasses already look remarkably similar to ordinary eyewear. Add facial recognition, and someone standing nearby may have no way of knowing whether they are simply being looked at or quietly identified.

Usually, you don’t want people poking around on your device. There are ways to stop this, including biometrics and passcodes, but Android users don’t have many privacy options once someone — whether it’s a kid, friend or thief — gets hold of their phone and gets past the lock screen.

Android 17 is potentially changing that with the introduction of a native app lock. The feature, spotted by 9to5Google, was introduced with the Android 17 QPR2 beta 3 build released late last week.

The feature, currently in testing, functions as an app lock should. You set it to lock an app, like a social media or messaging platform or photos archive. Then, anyone who wants to get into that app needs your passcode or biometrics (like face ID or a fingerprint) to access it. The app lock also disables notifications for that app and prevents any widget or shortcut from popping up on your home screen.

These functions prevent unauthorized users from snooping, making purchases, changing settings or accessing your personal data. 

Android’s app lock is managed in the settings under Security & Privacy, where you can enable or disable the feature for multiple apps at once. You can also enable it for one app at a time by long-pressing the specific app icon on your home screen, or locating it in the app drawer (menu).  

Google notes that “AI agents and services you’ve allowed” can still access data from the app, so if you want the app to be completely shut down, you’ll need to manually remove those permissions. 

Currently, the app lock feature is available only on Pixel devices running QPR2 beta 3. Google is launching the full version of QPR2 in December, when the feature will likely be expanded to more devices — such as Samsung — assuming that it makes it to full release. Google is known for testing features and then removing them entirely or saving them for future launches. That doesn’t seem likely in this case, but it’s still possible. 

Improving the security of your phone

The new built-in app lock feature would put Android on even footing with iOS, which introduced its app lock feature in iOS 18 back in 2024. It was possible to lock apps prior to this, but they required odd workarounds, such as creating an automation that triggered a screen lock when an app was opened.

Android 17 QPR2’s implementation is similar to iOS 18, where you lock an app by long-pressing its icon on your home screen, giving you the option to “require face ID” to unlock the app.  

Android users have been able to use privacy-based app locks for years through third-party apps or through features implemented by individual OEMs like OnePlus. These are strong enough to keep a casual friend from prying while they borrow your phone, but lack any true security features.

Android 17’s app lock should help shore up some of those security concerns and standardize the feature across Android devices — a welcome improvement for anyone seeking to deter unwanted attention.

Apple is quietly getting closer to launching camera-equipped AirPods, and the proof just turned up in a place nobody expected: a demo video tucked inside the macOS Tahoe 26.7 release candidate. MacRumors dug up the clip, and it’s the clearest look yet at how these AirPods might actually work.

What can the camera actually do?

In the short demo, a man holds up a book so the camera built into the AirPods can read the title. The voiceover explains that your world “becomes savable,” and all you have to do is ask to save whatever catches your eye. In simple terms, the camera feeds information straight to Visual Intelligence, and Siri can answer questions about your surroundings and log the details for you.

There’s also a nice little safety net built in. If your hair happens to cover the AirPods, you’ll get an alert telling you to clear it so the camera can actually see what’s in front of you.

When could we see these AirPods?

Internally, the camera AirPods go by the codename B790, and this isn’t the first time that name has popped up. Bloomberg’s Mark Gurman reported on the same product earlier and suggested a launch as soon as September. If that holds, we could see these AirPods debut alongside the iPhone 18 series, including iPhone 18 Pro and Pro Max, and the foldable iPhone Ultra.

Advertisement

And the AirPods aren’t the only surprise hiding in this update. macOS Tahoe 26.7 apparently references a long list of other unreleased Apple products too, making this release candidate a bit of a goldmine for leakers. Nothing is official yet, but between the leaked footage and Gurman’s earlier reporting, camera equipped AirPods are starting to feel less like a rumor and more like something Apple is getting ready to actually ship.

Relay, an AI-powered workflow automation tool that was launched in 2021 with the goal of becoming the new Zapier, is shutting down, and some of its staff — including its top executive — are joining Google’s Chrome team.

Jacob Bank, Relay’s founder and CEO, shared a company announcement Monday that reveals the app will be shutting down access for paying customers on September 14. Free customers would have already lost access as of August 15. The app’s closure was initially announced in July.

Bank, who previously spent a little over six years at Google, revealed that he would now be rejoining the tech giant as VP of Product for Google Chrome, where he will lead the product and developer relations teams for Chrome, according to his LinkedIn.

“I’ve spent my whole career doing one thing: building tools that help people get more done with AI, without sacrificing their personal creativity or insights,” said Bank in a post on X. “And joining the Chrome team is an ideal opportunity to bring those experiences to many, many more people.”

“We have some really ambitious plans to help you work with AI in Chrome to get things done, and I’ll have more to share soon,” he added.

Bank originally joined Google in 2015 after his previous startup, a scheduling app called Timeful, was acquired by the search giant. At Google, he worked in a number of different areas, including as product lead for Gmail, Google Calendar, and Google Chat, before eventually leaving to launch Relay.

Similar to Timeful, Relay offered users productivity upgrades via workflow automation, allowing businesses to streamline repetitive tasks like document drafting and copyediting, as well as various project management tasks. TechCrunch reached out to Google and to Relay for more information.

Just how Bank and others plan to further integrate AI into Chrome is a hanging question, though Bank calls Chrome “a perfect place to collaborate with agents.”

The move follows a steady path of AI integration into Google’s user experience. The integration of Gemini into search has fundamentally altered Google’s front end — not to mention the rest of the internet. Gemini has also been integrated into Chrome, where it serves as an optional in-browser assistant.

Overall, users seem to be responding well to these integrations, as Google recently reported that Gemini had crested 1 billion users.

The board base is very thick, which means you have to angle your wrists considerably to reach the tall keys. It’s in dire need of wrist support, which is an oversight since a lot of the best keyboards come with a rest. It’s also quite heavy, thanks to its metal construction. This does help to keep it planted, though, and it’s as solid as you would expect it to be.

To customize the inputs of the K8 Ultra 8K, you’ll have to use Keychron’s Web Launcher. Here you’ll find a reasonable selection of rebinding options, and there’s also a basic but functional SOCD function for altering key priority. It’s a relatively easy app to use, and despite being web-based, it’s quick and responsive; I didn’t encounter any disruptive loading times during my time with it.

Payments biz Stripe has reportedly finalized an agreement to acquire OpenRouter for at least $7 billion, signaling a shift in focus for the AI industry.

OpenRouter provides a way for customers to integrate multiple AI models into their workflow and infrastructure and is, by Ramp’s measure, the most popular of the gateway companies that have sprung up to simplify model integration. By acquiring the AI gateway service, Stripe would become a mediator of AI token sales, fitting its core business while gaining access to valuable data about AI model usage.

Stripe declined to comment, citing a policy of not addressing rumors or speculation. Prior to Bloomberg’s report about the agreement, the Wall Street Journal said acquisition talks were underway.

Akhil Verghese, founder and CEO of Krazimo, a developer of AI software for businesses, told The Register in an email that he found the acquisition fascinating because OpenRouter is to AI models what Stripe has been to financial infrastructure. 

“A key factor for me is the value,” he said. “I won’t pretend to be smart enough to understand how these valuations work, but $7 billion for OpenRouter surprised me. They raised at a $1.3B post money valuation in May, and even if revenue has grown significantly since then and acquisitions often have an additional multiple, 5x in 3 months is nuts.

“There are ways this acquisition makes sense, though. [Stripe CEO] Patrick Collison recently said that metered pricing is the native business model of the AI era. In purchasing OpenRouter, Stripe, which already controls where the money is going, now sees where the tokens are going.”

And there are a lot of tokens going around. An economics paper published last month, AI Premium, estimates total global LLM token consumption to be about 5 to 7 quadrillion per month, with about 2 percent of that being handled by OpenRouter.

While marquee model makers Anthropic and OpenAI have focused on building their own brands and steering customers toward their tooling, the reality is that business customers prefer options that avoid lock-in and maintain negotiating power. That expectation and the potential fees for gatekeeping have led to the proliferation of AI gateway firms and to $113 million in funding for OpenRouter back in May.

Verghese said rival AI gateways may find it more difficult to compete now that OpenRouter has access to so much funding.

Anthropic and OpenAI account for a substantial amount of the tokens processed by virtue of their subsidized flat-rate subscription plans – generally served directly rather than through a gateway or third-party provider. But as they push customers toward API pricing in preparation for going public, they risk driving business toward providers of more affordable, open weight models.

As of December 2025, OpenRouter reported serving more than 5 million developers to route traffic to more than 300 models from more than 70 providers. At the time, open weight models accounted for about 30 percent of tokens served. Today, the biz boasts more than 10 million developers and 80 providers.

OpenAI in October 2025 was serving about 8.6 trillion tokens per day, according to Andreessen Horowitz, and OpenRouter was serving more than 1 trillion tokens per day.

“The competitive frontier is no longer only about accuracy or benchmarks,” said Malika Aubakirova and Anjney Midha from Andreessen Horowitz at the time. “It is about orchestration, control, and a model’s ability to operate as a reliable agent.”

It’s about AI gateways and adjacent services that sit above models in the technical stack.

According to Vercel, which offers its own AI Gateway, open weight models have seen their share of gateway token spend grow from 11 percent in April to 36 percent in July. Meanwhile, the token spend collected by the four largest frontier labs, which had not dipped below 93 percent in seven months, fell to 89 percent last month.

Much of that came at Google’s expense, with Chinese AI labs Z.ai and Moonshot capturing the growth in open weight spending.

“GLM 5.2 and Kimi K3, both released in the last two months, are the first open-weight models running a meaningful share of the workloads historically owned by closed-weight labs,” Vercel said.

Even so, the top four frontier labs in the US continue to capture the lion’s share of spending. Vercel said those companies took 95 percent of spending through its AI Gateway in June. 

“In July, Anthropic collected 65 percent of gateway spending on 30 percent of token volume, at 4.4 times the average price of every other lab’s tokens,” Vercel said.

Open weight models, which have occupied the low end of the market (cheap tokens), have seen their fortunes shift, Vercel said. After their share of gateway token volume almost tripled between April and June, from 11 percent to 29 percent, capturing only four cents for every gateway dollar, they saw spending more than double to nearly nine cents for every dollar.

Open weight models, in other words, are becoming better and are bringing in more revenue even as the average token price declined 13.6 percent in July. If the trend continues and AI usage drifts further toward open weight models and away from proprietary frontier models, the AI gateway business should prosper.

Verghese expressed concern about whether OpenRouter will be able to maintain its neutrality under Stripe’s roof. The workload router isn’t supposed to play favorites in terms of where it directs tokens based on requirements, but should it start doing so, he said, the resulting loss of trust could be harmful. ®

Nothing’s budget-focused spinoff brand CMF has started teasing its next pair of wireless earbuds, and while the company is keeping the name to itself for now, there is a good chance we are looking at the successor to the CMF Buds Pro 2.

The teaser posted on CMF’s official X account shows a single in-ear earbud from the bottom. It has a metallic silver finish, a stem with what appear to be two microphone openings, and a silicone eartip finished in CMF’s signature orange. CMF simply calls it “a hint of what’s next,” without revealing a launch date or any specifications.

Why I think these could be the CMF Buds Pro 3

CMF could simply be preparing the regular Buds 3, but its previous release schedule makes the Pro model a strong possibility. The original CMF Buds Pro arrived in September 2023, several months before the regular CMF Buds launched in March 2024. CMF repeated the pattern with the next generation. Buds Pro 2 launched in July 2024, while the Buds 2 family followed in April 2025.

If CMF follows the same naming scheme, the teased earbuds would most likely be called CMF Buds Pro 3, rather than Buds 3 or any other 3 series model. There is still no confirmation, though. Other reports have suggested the teaser could simply point to the third-generation CMF Buds, so the exact model remains up in the air.

Advertisement

There is plenty CMF could improve

As a former CMF Buds Pro 2 user, I already have a pretty clear wishlist for whatever CMF is preparing next. The earbuds offered a lot for the money, including dual drivers, LDAC support, strong ANC, long battery life, and the Smart Dial built into the charging case.

The Smart Dial was easily one of my favorite features. Being able to adjust the volume or control playback without reaching for my phone was genuinely useful, but accidental presses did get annoying. I would also like to see CMF tone down the heavy bass and give users more control over the EQ. None of those are deal-breakers at the price point CMF usually targets, but they are areas where a new Pro model could make some meaningful improvements.

The biggest AI model release of the past few days, at least among the developers and AI power users on social media, wasn’t a frontier cloud model from OpenAI, Anthropic or Google.

It was a 27-billion-parameter model from Alibaba: Qwen3.8-27B landed on Hugging Face on Friday under an enterprise-friendly, open source Apache 2.0 license, giving developers downloadable weights for a dense multimodal model.

But Qwen3.8-27B isn’t a garden variety small local model: it includes native image and video understanding, a 262,144-token context window, configurable reasoning and support for coding and agentic workflows — a “compact, deployment-friendly” version of the capabilities developed for its Qwen3.8 generation.

That unusually small hardware footprint is a major part of Qwen3.8-27B’s appeal. Running the model at full 16-bit precision requires roughly 56GB of GPU memory, while an FP8 version needs about 28GB. But 4-bit quantization cuts the model itself to roughly 17GB, putting it within reach of high-end consumer machines such as a powerful gaming desktop or well-equipped laptop.

Hitting the sweet spot between capability and size

The outsized reaction among developers has been due to the dynamic combination of its capability and size.

Alibaba’s own launch benchmarks immediately supplied the first jolt. The company reported 61.7 on SWE-bench Pro, 90.3 on LiveCodeBench v6, 70.7 on its CoWorkBench office-work benchmark and 84.3 on OSWorld-Verified.

In Alibaba’s published comparison table, the 27B model even beats the listed Claude Opus 4.6 Max result on SWE-bench Pro and LiveCodeBench, although Opus remains ahead on Terminal-Bench, GPQA Diamond and Humanity’s Last Exam.

Some of Alibaba’s evaluations are internal, and benchmark harnesses are not identical across every comparison, making the numbers poor grounds for declaring a universal winner.

Third-party results show a powerful, local model with performance equivalent to proprietary models from months ago

The conversation changed Monday when third-party results began arriving.

Third-party AI benchmarking outfit Artificial Analysis gave Qwen3.8-27B a score of 52 on its Intelligence Index, a composite of nine evaluations spanning coding, science, reasoning and professional tasks. That happens to be the same score Artificial Analysis currently assigns OpenAI’s low-tier model GPT-5.6 Luna at its maximum reasoning setting — a proprietary offering only available over the cloud.

As open source coding agent Cline put it on X: “This is the first time a local model has scored frontier model capability. We weren’t expecting this pace of local progress anywhere near this soon.”

On Artificial Analysis’ Agentic Index measuring model performance on agentic tasks, meanwhile, Qwen3.8-27B scored 51, beating Claude Opus 4.8 on maximum reasoning effort — a frontier model Anthropic released less than three months ago.

That doesn’t mean these models are equivalent, but it helps explain why developers and AI power users stood up and took notice. As developer and AI podcaster/YouTuber Sero (@0xSero on X, real name Sharif Cherf) wrote on X: “A model that runs on 3k USD of hardware is beating everything from 4 months ago. Including Opus. Permanent underclass is cancelled.”

Developer Joshua “Xenova” Lochner, known for bringing machine-learning models into web browsers, highlighted the result Monday on X alongside an experiment running Qwen3.8-27B with custom WebGPU kernels. His reaction — “What a time to be alive!” — captures much of the mood: a model scoring in the vicinity of proprietary frontier systems can be downloaded, modified and executed locally rather than accessed only through a vendor API.

The appeal becomes clearer when the model is compressed. Developer and AI writer Simon Willison tested a roughly 17GB Q4_K_M quantization on an M5 Max MacBook Pro and Nvidia DGX Spark.

He found that it could write code, interpret images and operate a coding-agent loop through the Pi agent framework. In one experiment, the model navigated a codebase to explain how authentication worked; in another, it wrote and tested a Python utility Willison needed to convert an agent transcript from JSONL to Markdown.

“The fact that a 17GB file can do all of this stuff on my home machines is a miracle,” Willison wrote. His broader point is the one resonating with power users: capabilities that recently felt inseparable from expensive hosted models are moving into files small enough to keep on a workstation.

The reaction is showing up in usage as well. Cybernews reported Monday that Qwen3.8-27B passed 3 million Hugging Face downloads in its first three days, while quantized versions rapidly appeared for local inference tools.

The LocalLLaMA community on Reddit created a dedicated release megathread simply to consolidate the flood of benchmarks, quantizations, configuration advice and comparisons. One user showing a locally generated game described the model as “a different beast.”

Overthinking is an issue

That frenzy comes with an important caveat: Qwen3.8-27B appears to buy some of its quality by thinking a lot.

Artificial Analysis says the model generated 160 million output tokens across its Intelligence Index testing, versus a 43 million median for comparable open-weight models.

Willison encountered an extreme version of the same behavior because Qwen defaults to its xhigh reasoning setting. A request to generate an SVG of a pelican riding a bicycle took 21 minutes and consumed more than 22,000 reasoning tokens before producing the answer. He recommends starting with low or no reasoning for ordinary local use.

Investor and developer Tomasz Tunguz found a similar trade-off in a small nine-task test against DeepSeek V4 Flash: with reasoning enabled, Qwen edged ahead on quality in his agent stack, but he reported that it was roughly 30 times slower and 4.5 times more expensive. He explicitly cautioned that nine tasks were not enough for a verdict.

Inference software may narrow that gap. Qwen3.8-27B includes Multi-Token Prediction, and Willison reported about a 72% performance improvement on his DGX Spark after enabling MTP through llama.cpp compared with his default LM Studio configuration.

Even then, his normal LM Studio runs were producing only around 15 to 30 tokens per second — far below the responsiveness of many hosted models.

That tension is precisely why Qwen3.8-27B matters more than another leaderboard position.

What enterprises should take away from Qwen3.8-27B

For enterprises, the relevant comparison is not simply whether a 27B model “beats” Claude or GPT on a benchmark. It is whether a model small enough to run inside an organization’s own infrastructure can now perform enough coding, document analysis, vision and agent work to replace API calls for meaningful classes of tasks.

That proposition changes privacy, deployment and cost calculations. Apache 2.0 weights can be inspected, modified and hosted behind a company’s own controls, while Alibaba already documents compatibility with serving frameworks including vLLM, SGLang and TokenSpeed. Alibaba says a managed Qwen Cloud version with a 1-million-token default context and built-in tools is coming later.

The small size and accessible hardware requirements mean that enterprises, indie developers, and even curious consumers can easily deploy the model locally without worrying about their data leaving their machine — ensuring greater privacy, information security, governance and control.

There is a broader reason power users are paying attention. Hugging Face data reported by Business Insider this week shows that actual model usage skews dramatically toward smaller models even as enormous frontier releases dominate headlines; models above 70 billion parameters accounted for only a small share of 2026 downloads.

Alibaba’s strategy of publishing Qwen models across multiple practical size classes has helped make the family a recurring part of developers’ local deployment workflows.

Qwen3.8-27B pushes that logic further. Its benchmark scores still need more independent validation, its default reasoning behavior can be painfully inefficient, and no single leaderboard establishes frontier-model parity.

But three days after release, developers are no longer reacting primarily to Alibaba’s benchmark table. They are reacting to the experience of putting a comparatively small file on hardware they control and watching it perform tasks that, not long ago, seemed to belong exclusively to the largest proprietary systems.

For certain developers, AI power users—and yes, even enterprise deployments—that is the benchmark that matters most.