The AI industry is having a rogue agent summer. The latest model to escape onto the open internet during security testing is Kimi K3, a powerful open-weight offering from the Chinese company Moonshot AI.

Frontier Security, a US startup, says that Kimi K3 went outside of its sandbox while testing its defensive cybersecurity skills. As with incidents previously reported by OpenAI and Anthropic, the escape was partly enabled by a misconfiguration in the sandbox designed to contain it. Frontier claims, though, that the incident shows Kimi has fewer cyber safeguards than most other powerful AI models, something that allowed it to go off and use the internet without express permission.

“We found a leak in the sandbox,” says Yaron Singer, CEO of Frontier Security. “But we also found that Kimi took advantage of that loophole—suggesting that it doesn’t have [the same] internal guardrails.”

Unlike other recent incidents of AI agents going off-script, Kimi K3 did not hack anything after accessing the internet—because the answers to the problems it was seeking were easily attainable on GitHub.

Moonshot did not respond to a request for comment by time of publication.

The incident is the latest in a string of agent mishaps that suggest increasingly cyber-capable AI models are becoming more challenging to control.

Last month, OpenAI disclosed that an unreleased model had broken out onto the internet and then hacked Hugging Face, a company that hosts AI models and data, in order to find answers to problems it was tasked with solving. OpenAI subsequently shared that its AI agents had in fact hacked into four additional services as part of the spree.

Shortly after OpenAI reported its incident, Anthropic revealed that several of its models had also gained access to the internet and attacked outside systems. Last week, the AISI also disclosed that in its own testing, versions of OpenAI and Anthropic models that had security safeguards disabled perpetrated multiple hacks across the internet, including a particularly ambitious attempt by Anthropic’s Mythos 5 to plant malicious code in an open-source project on GitHub.

While these AI hacking episodes all vary in both cause and degree, the Kimi K3 is similar to several of them in that a misconfigured sandbox allowed access to a number of websites rather than keeping it contained to a simulated environment. The model was expressly tasked with solving problems that should not have involved going off to find the answers online, and appears to have gone outside of those instructions. The model had to figure out for itself that it had access to certain websites by probing the network settings of the sandbox.

While human error appears to have played a major role in each of the breakouts, the consequences have been compounded by the fact that advanced AI models are designed to use reason and take complex actions in order to solve problems.

Another key difference between previous incidents and the one discovered by Frontier Security is that it involves a model that is already widely available, with the same safeguards an average user would encounter.

“Kimi K3 is very good at following a goal by any means necessary and also doesn’t have the guardrails to prevent it from cheating or escaping the sandbox,” says Paul Kassianik, a researcher at Frontier Security.

Kassianik and Singer both say that Kimi and other open-weight models are also excellent tools for cybersecurity defense. (Hugging Face ultimately used an unnamed AI model from China to defend itself against the OpenAI agent hack.) Their company has developed benchmarks that measure a model’s capacity to find vulnerabilities in software and networks, which show that Kimi excels at these tasks.

Apple can usually make suppliers bend the knee. It buys components in volumes few consumer technology brands can match and has built a supply chain that rivals spend decades trying to copy. However, that leverage has now run into a memory manufacturer willing to say no.

Apple reportedly approached China’s CXMT as an additional DRAM supplier and asked for lower prices. CXMT instead quoted figures comparable to or even higher than Samsung and SK Hynix. Huawei, Xiaomi, and other Chinese manufacturers have already locked up much of its output through higher-priced long-term deals, leaving the company with little reason to accept Apple’s terms.

I felt a much smaller version of the same crisis recently. A failing Corsair Vengeance LPX 8GB DDR4-3200 stick made my aging PC unstable, so I had little choice but to replace it. The same module should have cost around 1,500 INR, or roughly $16, before the RAM crisis. I paid 8,000 INR, around $84.

Paying more than five times the sensible price for old DDR4 hardware is absurd. Apple has billions of dollars and purchasing power that I obviously do not. Yet in today’s market, even Apple is learning that scale may secure memory supply without necessarily making it affordable.

Apple’s supply-chain muscle has finally hit a wall

CXMT’s growing influence should eventually be good for the memory industry. It gives device makers another serious supplier beyond Samsung, SK Hynix, and Micron, potentially reducing their dependence on the three companies that have dominated DRAM for years.

More competition usually leads to better pricing. Unfortunately, CXMT is expanding into a market where nearly every available memory chip already has a buyer waiting for it. Samsung and SK Hynix are investing heavily in high-bandwidth memory for AI accelerators, where the profit margins are far higher than those of ordinary consumer DRAM. Chinese smartphone and computer manufacturers are securing CXMT’s conventional memory before it leaves the factory. Apple cannot use one supplier to pressure another when all of them are already struggling to meet demand.

The company has reportedly asked the US government to let it source DRAM from CXMT and NAND storage from YMTC for products sold outside the country. Approval would likely help Apple secure enough memory, but it may do little to bring down what the company pays. CXMT has already shown that it does not intend to become Apple’s bargain supplier, weakening the company’s hopes of using another manufacturer to rein in rising memory costs.

Memory costs are already reaching Apple buyers

Apple spent months shielding customers from rising component costs, but it has now started passing them on to buyers. Prices for several Macs and iPads have already gone up by hundreds of dollars, and the iPhone 18 Pro lineup may be next.

TrendForce says conventional DRAM contract prices jumped between 93% and 98% in the first quarter of 2026, pushing industry revenue up 81% to $97 billion. By July, prices were still rising, although the pace of increase had started to slow. Counterpoint Research estimated the 12GB and 1TB iPhone 18 Pro Max could cost Apple nearly $300 more to build than its predecessor, with memory and storage responsible for the largest part of that increase.

Supply is not growing fast enough to bring prices back down. IDC expects the amount of DRAM available this year to rise by only 16%, while industry revenue is forecast to jump 177% to $418.6 billion.

Apple can absorb the increase, accept lower margins, reduce memory or storage configurations, or pass the cost on to buyers. We have already seen which option it picked for Macs and iPads. The iPhone may not remain protected for much longer.

The blame game is in full swing, but it won’t lower your bill

Micron has its own explanation for the current mess. The company claims a couple of major customers used their purchasing power during the previous memory downturn to secure rock-bottom prices. Those prices supposedly made expansion difficult to justify, leaving manufacturers unprepared when demand returned. Micron did not publicly identify Apple, although reports have connected the remarks to the company.

I find Micron’s argument a little ridiculous.

A company placing massive, predictable orders will always expect a volume discount. Apple does not buy DRAM at retail, and Micron, Samsung and SK Hynix were not producing millions of components for one of the world’s most profitable companies as a charitable exercise. I also find it difficult to accept that Micron continued serving Apple for years without earning enough money to keep the relationship viable. If the contracts were genuinely unsustainable, Micron had the option to renegotiate them or sell its production elsewhere.

Memory is one of the most cyclical parts of the technology industry. Manufacturers expand production when prices are high, excess supply pushes prices down, investment slows, and the next shortage sends them back up. Blaming customers for negotiating during the low part of that cycle feels awfully convenient now that suppliers have regained pricing power.

Micron does not have a spotless history here, either. A former Micron sales manager pleaded guilty to obstructing the US Justice Department’s DRAM price-fixing investigation after withholding and altering subpoenaed documents. Samsung and Hynix later pleaded guilty to participating in the conspiracy, paying fines of $300 million and $185 million, respectively. None of that proves anything illegal is happening today, but it gives me little reason to accept its version without question.

Still, Micron is right about one part of the crisis. AI companies are buying huge quantities of highly profitable HBM, and memory manufacturers have every reason to prioritize those customers. Conventional RAM for phones, laptops, consoles, and PC repairs is left competing for what remains.

CXMT’s rise could eventually loosen the grip of the established memory manufacturers, but more competition needs more production before buyers see any relief. For now, the new player has entered the same expensive market rather than disrupted it.

Apple will pay more, device makers will pass those costs along, and customers will delay upgrades or accept higher prices. Others, like me, will pay $84 for an 8GB DDR4 stick because an aging PC suddenly leaves them with no other choice.

The Shards is guaranteed to have audiences talking, for better or for worse, as Bret Easton Ellis’ controversial semi-autobiography comes to the small screen.

But, in my opinion, Ryan Murphy teaming up with Ellis has given him the chance to shine once again. Fans of Murphy’s work were left disappointed and baffled in recent months after his abysmal Hulu series All’s Fair aired, which has a seriously low Rotten Tomatoes score of just 6%.

PC gaming sites and social media discussions were buzzing this week after a report from a Chinese forum claimed PC motherboard prices could soon rise significantly.

Websites including VideoCardz, Guru3D and Digital Foundry, among others, cited what appears to be a supply chain report on a Chinese forum called Boards Channels, suggesting that PC motherboard makers could raise prices for consumers as material costs climb by as much as 50% in the coming months.

The link to the Board Channels report no longer seems to work, but a screenshot posted on these sites suggests that Taiwanese motherboard manufacturers, including ASUS, Gigabyte and MSI, will be trying to address the effects of rising component costs.

Whether the report is legitimate or not, we have some bad news, especially if you plan to build a PC or perform an upgrade. PC gamers, builders and upgraders have already seen the price of hard drives and SSDs skyrocket following massive cost jumps for RAM (dubbed by some as RAMageddon) over the last year due to AI data center needs and component shortages. Experts consulted by CNET said the implications of the Board Channels report for motherboard pricing are not only credible but also highly likely to materialize.

Why these price increases could happen

One reason analysts believe motherboards are likely to go up in price is that practically all electronic components are being affected by the AI data boom, with data centers gobbling up memory and storage products, creating supply constraints. There is also a reduced supply of materials such as copper, as well as chemicals manufactured in parts of the Middle East, said Cassandra Cummings, CEO of the electronics design and manufacturing company Thomas Instrumentation Inc.

“You can’t make PC motherboards without memory chips or copper, so its base pricing will follow what’s happening in those markets,” Cummings said in an email.

The war in Iran is also hurting supply chains.

“Some of those factories have been damaged in the war, and shipping logistics are struggling, so I believe those chemicals are also in short supply these days, causing more cost increases for the industry,” Cummings said.

Gaming enthusiasts may also be competing for motherboard parts with big PC makers, such as Dell and Apple, which could have a lock on certain supplies and components from some of the same motherboard manufacturers.

Joshua Goshorn, senior director at Quandary Peak Research, said in an email that he found the Board Channels report credible.

“With so much demand generally for PCBs (printed circuit boards) and materials and processes needed to create the PC motherboard, it seems reasonable that the cost of the PC motherboard from MSI, ASUS, etc. will increase,” he said.

But Goshorn believes those cost increases could be modest compared to the increases the market has seen in RAM and hard drive prices.

Companies that make PCs have an advantage over solo builders and upgraders because they have access to volume purchasing, and some of them have stockpiles of components ahead of any price shifts.

Although he hears that production of circuit boards is continuing to ramp up, that may not mitigate the rising cost of materials.

“I don’t believe that pricing relief to the consumer will come any time soon,” Goshorn said.

Calm before the storm

PC component costs don’t just affect DIY PC owners; they also create cost headaches for companies that make pre-built PCs for gamers and power users. The head of one of those custom builders, Maingear, says that while we haven’t seen a cost jump for motherboards just yet, it could be coming.

“Motherboards have stayed relatively flat” compared to storage and memory,” William Santos, CEO at Maingear, said in an email. “If you’re looking into upgrading your PC, we recommend acting sooner rather than later to avoid the next component cost increase. Waiting doesn’t seem to pay off in this market.”

AI and ML

AI usage is evident but isn’t yet a serious problem

The 35th USENIX Security Symposium (USS), which takes place next week in Baltimore, Maryland, hit an all-time high for paper submissions.

While some of that increase has been aided by the availability of AI tools, those managing the conference say abuses were minimal due to defensive measures. But they’re also trying not to look too closely in order to preserve trust within the security research community. 

“This year’s conference has received ~3,030 valid submissions (~1,280 in Cycle 1 and ~1,750 in Cycle 2),” explained Ben Stock, tenured faculty at the CISPA Helmholtz Center for Information Security and USS program co-chair, in an email to The Register. “This is up from the previous year, which had ~2,400 submissions in total.”

Stock said that the entire security community has seen growth of this sort and  pointed to the Network and Distributed System Security Symposium (NDSS),  which saw its paper submission count jump from 694 in 2024 to 1,311 in 2025 and 1,481 this year.

“So, I would not call the growth unprecedented, even though the number of submissions has reached a high point compared to previous years,” he said. “This is something we had expected and scaled our Program Committee (PC) accordingly.”

Sussing out unacceptable uses of AI

A paper published in April, “More Versus Better: Artificial Intelligence, Incentives, and the Emerging Crisis in Peer Review,” found that since the release of ChatGPT in 2022, submission volume at major academic journals has increased 42 percent.

In the USENIX Security ’26 transparency report, issued in January between the first and second paper submission cycles, Stock and fellow co-chair Elissa Redmiles, assistant professor of computer science at Georgetown University, detail how they’ve developed tools and policies to account for the possibility of AI usage, both for paper submissions and in paper reviews.

“The proliferation of readily-available LLMs to aid in writing and developing code is not unknown to the community,” their report says. “However, we see an alarming trend of AI usage in key areas of the scientific process. Therefore, we took actions against two types of identifiable actions which violate the scientific process in our minds: non-existing (possibly hallucinated) references and usage of AI in the review process.”

After identifying and rejecting a paper that contained nonexistent references, the report explains, the conference organizers developed tooling “to extract references from the submitted PDFs, query well-known sources such as DBLP and arXiv, and manually confirm invalid references.”

The org rejected papers containing three or more hallucinated references, a policy that impacted 21 of the 1,181 first round submissions (1.78 percent).

“We have rejected papers for the repeated presence of nonexistent references,” said Stock. “We cannot say with certainty that these were AI-hallucinated, but nevertheless considered these papers to be problematic and thus rejected them.”

The report notes that more than 100 additional papers contained at least one reference that reviewers could not confirm. Aware that some of these might simply be false positives due to name spelling differences or missing citations, conference officials opted not to investigate these in order not to further burden staff.

Conference organizers draw the line at using AI for bibliography preparation. “We believe that it is critical to halt this trend that threatens scientific integrity before it grows further,” the report states.

However, limited use of AI to polish human-written text is expected, and that extends to those reviewing submitted papers, up to a point.

“We have not set a dedicated AI policy, but have made it clear to our PC members that usage of [AI] services to write reviews is not permitted, in particular also because this violates confidentiality,” said Stock.

“We have detected a tiny number of cases where we have reached sufficient confidence that AI was used and took appropriate actions, including removal of the members from the PC and allowing affected authors to resubmit.”

Under that policy, USS asked five of 496 reviewers to cease participation.

“We have not seen evidence that leads us to believe that AI generated submissions have become a significant challenge for the security community,” said Stock. “This does not mean that AI hasn’t been used in parts of these submissions, though.” ®

The programme aims to teach professionals to use generative AI tools effectively, safely and responsibly at work.

A new, fully funded course designed to help professionals develop their practical AI prompt engineering skills for use within their organisations is to be offered by CeADAR, Ireland’s national centre for applied AI. 

Developed by CeADAR’s European Digital Innovation Hub (EDIH), the Prompt Engineering for You programme aims to teach professionals to use generative AI tools effectively, safely and responsibly in real-world settings. 

Funded by the European Commission and Enterprise Ireland, and aligning with the Department of Enterprise, Tourism and Employment’s ‘AI – Good for Business’ initiative, the course will teach hands-on skills via five modules that explore what prompt engineering is, relevant techniques, and how to design clear and effective prompts.

The course can be engaged with flexibly, is aimed at professionals at any stage of their AI journey, and does not require a technical or coding background. Once the course has been finished, participants will receive a certificate of completion from CeADAR’s EDIH for AI programme.

CeADAR’s director of innovation, development and EDIH for AI, Ricardo Simon Carbajo, said, “With so much noise surrounding AI, it can be difficult for people and organisations to know where to turn for trusted guidance. 

“At CeADAR, our goal is to cut through that clutter and help organisations navigate as they adopt AI. Building on the success of our AI for You course, CeADAR is now releasing Prompt Engineering for You to continue equipping organisations to confidently live and work alongside these technologies.”

The announcement of the new course comes at the same time as a report showing that Ireland remains one of the EU’s strongest digital performers. 

The European Commission’s ‘2026 Digital Decade Country Report for Ireland’ indicated that the country now ranks second in the EU for basic digital skills, with 83pc of the population possessing at least basic digital skills, compared with an EU average of 60pc.

The report also ranked Ireland fifth in the EU for generative AI adoption, at 45pc, significantly above the EU average of 33pc.

Commenting on the report, Minister for Enterprise, Tourism and Employment Peter Burke, TD said, “The Digital Decade report confirms Ireland’s position as one of Europe’s leading digital economies. Our strong performance reflects sustained investment in digital infrastructure, skills and innovation. 

“As Ireland holds the Presidency of the Council of the European Union, we welcome the progress being made across Europe towards the Digital Decade 2030 targets and look forward to working with our European partners to strengthen Europe’s digital competitiveness and AI ambitions.”

Minister of State for Trade Promotion, AI and Digital Transformation Niamh Smyth, TD added, “Ireland’s strong performance reflects the progress we are making through sustained investment in skills, connectivity and innovation. It is particularly encouraging to see strong levels of AI adoption among Irish businesses.”

Don’t miss out on the knowledge you need to succeed. Sign up for the Daily Brief, Silicon Republic’s digest of need-to-know sci-tech news.

The wireless headphone market has become a knife fight, with Sony, Apple, Sennheiser, Bowers & Wilkins, and Bose all competing for the same premium buyers. The newly announced Bose QuietComfort Headphones 2nd Gen enter that battle by borrowing one of the QuietComfort Ultra 2nd Gen’s biggest features, further blurring the line between Bose’s two flagship noise-cancelling models.

Built around the company’s long-running noise-reduction platform, the QC Gen 2 combines a refreshed industrial design with Bose TrueSpatial technology and three Immersive Audio modes: Still, Motion, and Cinema. The result is a more modern version of one of the most recognizable wireless headphones on the market, although Bose may now have to explain why some buyers should still pay more for the Ultra.

Related Reviews:

Immersive Audio Is No Longer Reserved for the Ultra

Bose TrueSpatial technology is the most significant upgrade, bringing three Immersive Audio modes—Still, Motion, and Cinema—to the standard QuietComfort line for the first time.

Unlike spatial-audio systems that require content mixed in Dolby Atmos or another immersive format, Bose Immersive Audio can process conventional stereo from virtually any source. Its digital signal processing moves the presentation outside the listener’s head and creates the impression of listening to a pair of stereo loudspeakers positioned in front of them.

Still Mode: Anchors the virtual soundstage in front of the listener. As the listener turns their head, the apparent position of the music remains fixed in the room, similar to listening to stationary loudspeakers.

Motion Mode: Keeps the virtual soundstage aligned with the listener’s head as they move. It is intended for walking, commuting, and other situations in which a room-anchored presentation could become distracting.

Cinema Mode: Expands the perceived soundstage for movies, television, and spoken-word content while keeping dialogue focused in the center. Background effects are distributed more broadly to create a larger, more theatrical presentation without requiring a native surround or spatial-audio soundtrack.

The important distinction is that Bose is not adding more channels to the original recording. TrueSpatial uses processing and head tracking to reinterpret two-channel audio as a wider, externalized listening experience.

Sound Design

The QuietComfort Gen 2 retains Bose’s proprietary SoundDesign digital signal processing, but adds lossless wired playback over USB-C at up to 24-bit/48 kHz. The distinction matters: lossless audio is available through the USB-C connection, not over Bluetooth.

Noise Cancellation That Adapts to the Fit

Bose has renamed its noise-cancellation platform QuietControl, but the QuietComfort Gen 2 also introduces several meaningful refinements over the first-generation model.

New adaptive feedforward controls are designed to improve performance when the earcups cannot form a perfect seal, which can happen when listeners wear glasses, hats, or have hair trapped beneath the cushions.

The six-microphone system can now better compensate for those small gaps, helping the headphones maintain more consistent noise cancellation under less-than-ideal fit conditions.

Bose has also refined ActiveSense in Aware Mode. The system applies noise cancellation more smoothly and naturally when sudden sounds occur, reducing sharp spikes from passing trains, traffic, or other loud interruptions without completely shutting out the listener’s surroundings.

For a more traditional listening experience, noise cancellation can be adjusted manually or switched off entirely through the Bose app for Android and iOS.

Listeners can also create Custom Modes that combine their preferred levels of noise cancellation and acoustic transparency. These settings can be cycled quickly using the action button on the left earcup.

Refined Comfort and Style

The QuietComfort Gen 2 features a redesigned headband with smoother adjustable sliders and softer synthetic leather cushioning that rests more comfortably against the head.

Bose has also optimized the clamping force to provide a secure fit without making the headphones feel heavy or overly tight. Newly shaped oval earcups are designed to follow the natural contours of the listener’s head and improve long-term comfort.

Contrasting color accents inside the earcups add a subtle two-tone look that gives the familiar QuietComfort design a more modern appearance.

New Colors

Alongside the standard Black and White Smoke finishes, Bose is offering three limited-edition color options:

• Eucalyptus Green: A muted, nature-inspired green with warm yellow undertones, designed to complement current fashion and interior color trends.
• Dewdrop Mint: A playful 1990s-inspired finish that combines a soft mint exterior with teal accents inside the earcups.
• Rosewood Mauve: A deeper, more expressive option that pairs a rich mauve finish with vibrant fuchsia accents for a bold two-tone appearance. 

Controls

The QuietComfort Gen 2 uses physical buttons positioned on both earcups, allowing listeners to control playback, calls, and listening modes without reaching for their phone.

Buttons on the right earcup handle primary functions, including play/pause, volume adjustment, and answering or ending calls. A dedicated button on the left earcup cycles through listening modes and can be assigned as a shortcut for the connected device’s voice assistant or Spotify Tap.

Spotify Tap should not be confused with Spinal Tap, although Bose has yet to confirm whether the volume control goes to 11.

Connectivity

Wireless connectivity for the QuietComfort Gen 2 is provided by Bluetooth Core 5.4 with multipoint connectivity that allows seamless switching between two connected devices. This makes it easy to move between music, calls, and more. Android users can also take advantage of Google Fast Pair for simplified setup.

Two-way USB-C audio supports both high-quality wired playback and voice input through the headphones’ microphones, making the QuietComfort Gen 2 suitable for calls and videoconferencing through apps such as Zoom and Microsoft Teams.

Bose also includes a USB-C-to-3.5 mm cable for connecting the headphones to analog sources, including seatback entertainment systems on aircraft.

Battery Life

The QuietComfort Gen 2 provides up to 24 hours of listening time, dropping to 18 hours when Bose Immersive Audio is enabled. That places it ahead of the AirPods Max 2, which delivers up to 20 hours with Active Noise Cancellation, but behind the Sony WH-1000XM6 and its 30-hour rating with noise cancellation switched on. Apple’s 20-hour estimate includes Spatial Audio, however, giving it a slight advantage when Bose’s comparable Immersive Audio processing is active.

Bose allows the headphones to continue playing while they are being charged, so a depleted battery does not have to end a wired listening session. A 15-minute quick charge provides up to 2.5 hours of additional playback.

Bose App

The Bose app for Android and iOS provides a three-band equalizer, customizable listening modes, and configurable shortcut controls. Users can also manage connected Bluetooth devices, install firmware updates, and adjust additional headphone settings from one central interface.

Comparison

The Bottom Line 

Bose built its reputation on noise-reducing headphones, and the QuietComfort 2nd Gen continues that tradition. It also enters one of the most competitive segments of the wireless headphone market, facing the Sony WH-1000XM6, Apple AirPods Max 2, Sennheiser MOMENTUM 5 Wireless, and Bowers & Wilkins Px7 S3 at nearby premium price points.

Bose and Sony remain the strongest choices for buyers who prioritize noise cancellation and travel comfort, while AirPods Max 2 will have considerable appeal for listeners already invested in Apple’s ecosystem. Those who place sound quality ahead of maximum noise isolation should also audition the Px7 S3 and MOMENTUM 5 Wireless, both of which offer compelling alternatives without moving into a substantially higher price category.

That brings us to the $450 question that will matter most to potential buyers: if the QuietComfort Gen 2 now includes Bose Immersive Audio, USB-C lossless playback, and many of the features that once separated the two models, does spending more on the flagship still make sense?

Bose has not announced a QuietComfort Ultra 3rd Gen, but the shrinking feature gap raises an obvious question: how long can the current Ultra remain sufficiently different to justify its higher price?

For now, the QuietComfort Ultra 2nd Gen remains Bose’s flagship over-ear model and retains several meaningful advantages:

• A more premium design with polished metal components and softer cushioning
• Bose’s most advanced over-ear noise-cancellation performance
• CustomTune technology, which calibrates the sound and noise cancellation to the listener’s ears
• On-head detection with automatic Bluetooth standby and a low-energy deep-sleep mode
• A capacitive touch strip for volume adjustment
• Snapdragon Sound support, including lossless wireless playback with compatible Android devices

Those extras make the Ultra the stronger choice for frequent travelers, Android users with compatible Snapdragon Sound devices, and buyers who want the best noise cancellation Bose currently offers.

For everyone else, the QuietComfort Gen 2 may now represent the better value. It delivers the core Bose experience, including Immersive Audio, physical controls, USB-C audio, and adaptive noise cancellation, without charging buyers for every premium feature in the cabinet.

That leaves Bose with a positioning problem. The QuietComfort Ultra remains more luxurious and technically complete, but it no longer feels like an entirely different class of headphone. Whether its stronger noise cancellation, CustomTune processing, premium construction, and lossless wireless support are enough to justify the higher price will ultimately depend on real-world performance.

eCoustics Headphone Editor Will Jennings is currently testing our review pair, so we should have answers to those questions rather soon.

Price & Availability

The new Bose QuietComfort Gen 2 Headphones will open for preorder beginning August 8 on Bose.com and through select resellers for $359, and will start shipping on August 13. Available in Black, White Smoke,  or three new limited-edition colors: Eucalyptus Green, Dewdrop Mint, and Rosewood Mauve. 

The previous Bose QuietComfort headphones are priced at $359 at Amazon, but are likely to go on sale as phased out.

The Bose QuietComfort Ultra Gen 2 headphones are priced at $449 at Amazon

Related Reading:

Earlier this week, the AI startup Liquid, formed in 2023 by former MIT computer scientists, debuted LFM2.5-2.6B, a new open-weight language model designed specifically for agentic workloads.

In release materials and a recent interview with VentureBeat, Liquid’s researchers said LFM2.5-2.6B can run entirely on local hardware — from smartphones and laptops down to a Raspberry Pi — without relying on cloud inference or GPUs, unlocking edge AI applications and giving more options to enterprises working in regulated industries or with sensitive information they don’t want to send up to the cloud.

It’s best suited for high-volume, well-defined agentic tasks that run locally — tool calling, document management, calendar and workflow automation, and always-on background routines — and for connectivity-limited environments like vehicles and robotics, though coding-heavy work is better left to larger models.

Even for those businesses without such concerns, the appeal of running performant, task-specific agents at the cost of essentially electricity, may be enough to make the new model quite appealing.

But the custom open weights license, as with Moonshot’s larger frontier model Kimi K3 released last month, is worth a close look by enterprise legal teams.

The basics

LFM2.5-2.6B contains 2.6 billion parameters, supports a 128,000-token context window, and includes native tool calling. The somewhat tricky name is explained by the generation of model (2.5) combined with the parameter count (2.6B).

Both the post-trained model and a base checkpoint (LFM2.5-2.6B-Base) for developers who want to fine-tune it are available now on Hugging Face, with day-one support for major inference stacks including llama.cpp, MLX, vLLM, SGLang, and ONNX — positioning it for deployment across consumer hardware, enterprise infrastructure, and embedded systems.

Liquid also offers an open source fine-tuning framework, LEAP.

Rather than positioning LFM2.5-2.6B as a competitor to the largest frontier models, the company is making a different argument: that a sufficiently capable small model can unlock categories of enterprise applications where latency, privacy, deployment flexibility, or inference costs matter more than absolute benchmark leadership.

“I do also believe that the best models will be in the cloud, and there’s no problem with that,” Maxime Labonne, Liquid AI’s head of post-training, told VentureBeat in an interview following the launch. “We want to make models for another type of user, and the best way of describing it is: you should use [edge AI] when you can’t use a cloud model.”

Small enough for a Raspberry Pi

Asked about the minimum viable hardware, Labonne said the model runs “very, very well” on CPUs — and that the LFM2 architecture underlying the model was explicitly designed around real-world CPU performance rather than GPU benchmarks.

“I think the best example is a Raspberry Pi,” he said. “We have a lot of demos that show that actually, it works pretty fast on the Raspberry Pi.”

Company-reported measurements indicate decoding throughput of approximately 220 tokens per second on an Apple M5 Max and 113 tokens per second on an AMD Ryzen AI Max+ 395, while using less than 2.5 GB of memory — and around 30 tokens per second on a smartphone. Users can try the models on their phones through Apollo, Liquid AI’s mobile app.

At the other end of the deployment spectrum, Liquid AI reports the model reaches nearly 15,000 output tokens per second on a single Nvidia H100 GPU under sustained concurrent load — roughly 1.3 billion tokens per day on one card. These figures are vendor benchmarks and have not been independently verified.

For Labonne, memory footprint and speed are not conveniences but hard constraints that determine what can be deployed at all.

“What we want to show is that it’s a really good trade-off, because you get the level of quality that you get with much bigger models, but in a tiny, tiny form factor,” he said. “You can deploy it in target devices where you are not able to deploy the other ones at all.”

Trained for agents instead of chatbots

Liquid AI says LFM2.5-2.6B was developed around the assumption that language models are increasingly consumed through agent frameworks rather than traditional conversational interfaces.

“Models are not consumed in chatbots anymore. They’re really consumed through agentic harnesses, like OpenClaw, like Hermes Agent,” Labonne said. “We wanted to make sure that this model is not just good at math or at code, but it’s good at using tools.”

The model is pretrained on approximately 34 trillion tokens, with a vocabulary doubled to 128K to better support non-Latin scripts and a dedicated mid-training phase to extend the context window to 128K tokens for long-running agent workflows.

Post-training follows a four-stage pipeline: supervised fine-tuning, teacher specialization (training separate expert models for domains like instruction following, math, code, and tool use), multi-domain on-policy distillation (MOPD) to merge those experts’ capabilities back into a single student model, and finally agentic reinforcement learning.

During that last stage, the model was trained directly inside production agent harnesses — including Hermes Agent and OpenClaw — on realistic productivity tasks involving research, coding, document management, tool invocation, and workflow automation, exposing it to those harnesses’ actual tools, system prompts, and interaction patterns.

Labonne described the pipeline overhaul as producing a “happy accident”: gains that extended well beyond the agentic targets.

“Through these new training techniques, we also got a lot better at everything. We got better at math, at instruction following. We’ve never been good at code, actually — and with this, we even got really good at code,” he said.

Building the model — and the harness

Notably, Liquid AI also built its own agent harness rather than relying solely on existing frameworks, and demonstrated the model running inside it on a phone, planning and calling tools entirely on-device.

“This is a harness running on a phone, and I don’t know if there’s any other harness running on a phone,” Labonne said.

The company had two reasons, he explained. The first was necessity — no phone-native harness existed. The second is a different interaction model: today’s harnesses wait for a prompt, and Liquid AI wants assistants that act on their own.

“We want proactive agents. We want agents that run in the background, check what you’re doing, check your calendar, and based on this context, do tasks,” he said. “That doesn’t exist today, really.”

Co-designing the harness and model also lets the software compensate for the model’s weak spots. “Everything that the model is bad at, the harness should help the model with — provide as much assistance as possible to make it more reliable,” Labonne said. “End users don’t care if it’s the model or the harness. What they want is that the task is achieved at the end of the day.”

The model nevertheless works out of the box with established harnesses including Hermes Agent, OpenClaw, and Pi, served behind any OpenAI-compatible endpoint.

Swap the harness, not the model

For enterprise deployment, Labonne argued the release marks a shift in what small models can be used for. Until now, he said, local models made economic sense mainly as narrowly fine-tuned specialists — trained to do one thing at cloud-model quality, much faster and cheaper. Agentic capability changes that calculus, because the same model can be repurposed by changing the tools around it rather than the model itself.

“You can have a calendar assistant, and you can reuse the same model and make a meeting assistant that will record what everybody said and summarize it — a bit like Granola, for example,” he said. “You don’t change the model; you just change the harness. You just change the tools around it. This gives much more generalizability, and it’s a lot easier to do and a lot cheaper as well.”

He still recommends fine-tuning for production deployments whenever feasible: “If you don’t fine-tune it, you leave some quality on the table. If you fine-tune it well, it’s going to match the performance of GPT and Claude — really, if your task is not the most complex task in the world,” he said, adding that the barrier to entry has collapsed: “The bar to be able to do fine-tuning now is super low. It’s very accessible to everyone.”

How it stacks up against DeepSeek-V4-Flash, Google’s Gemma and Alibaba’s Qwen

Liquid AI released its own benchmark comparison charts pitting LFM2.5-2.6B against the models enterprises are most likely to shortlist for the same edge deployments: Google’s Gemma 4 E2B (5.1B parameters) and E4B (8B), and Alibaba’s Qwen3.5-4B (4.7B) and Qwen3.5-9B (9.7B).

A separate test by local AI client platform Atomic Chat found that LFM2.5-2.6B completed 35 tool calls to complete three tasks (checking weather and local time in six cities, converting one budget into six currencies, checking four hotels and booking for a date) 3.7 times faster than DeepSeek-V4-Flash (a whopping 284B parameters), the model has skyrocketed to the top of OpenRouter since its release last week.

Gemma 4’s small models are multimodal generalists, accepting image and audio input alongside text, and use a Per-Layer Embeddings design that keeps only a fraction of their weights active per token — which is why Google markets them by “effective” size (2.3B and 4.5B) despite total footprints of 5.1B and 8B. Alibaba’s Qwen3.5 small series, released in March, is natively multimodal from 4B up and leans on scaled reinforcement learning to chase frontier-style reasoning — Alibaba touts the 9B model as matching or beating OpenAI’s far larger gpt-oss-120B on reasoning benchmarks.

LFM2.5-2.6B takes a narrower path: it is text-only, dense, and specialized for agentic work, with Liquid AI shipping separate vision and audio variants of the LFM family rather than folding everything into one checkpoint.

Where Qwen’s post-training reinforcement learning targets reasoning, Liquid’s targets tool use inside real agent harnesses.

The result, per the company’s published numbers, is that the smallest model in the comparison leads every instruction-following benchmark (IFBench, Multi-IF, IFStruct) and nearly every tool-use benchmark — 77.83 on ToolSandbox versus 76.44 for Qwen3.5-9B, a model nearly four times its size — trailing only that 9B model on BFCLv4.

On agentic evaluations it beats both Gemma models across the board and essentially ties the Qwens: 26.89 on BrowseComp+ versus 27.23 for Qwen3.5-9B. It also posts the best score on AA Omniscience, a knowledge benchmark that penalizes hallucination.

The Qwen models keep the edge where their training focus lies: math (Qwen3.5-9B leads AIME25) and coding, where larger models retain an advantage on LiveCodeBench — though Labonne noted the gap is smaller than the parameter counts would suggest.

“With LiveCodeBench v6, we might not be the best among these models, but we’re also by far the smallest. Showing that we’re competitive with them is already quite a big win for me,” he said.

One differentiator cuts the other way: licensing. Gemma 4 and Qwen3.5 ship under the permissive Apache 2.0 license — a change Google made specifically to court enterprises. DeepSeek-V4-Flash ships under a similarly permissive MIT License.

Meanwhile, Liquid AI’s revenue-gated license (detailed below) asks larger companies to strike a commercial deal. Enterprises above the threshold are effectively trading license friction for footprint and tool-use performance.

Licensing reflects a commercial middle ground

LFM2.5-2.6B is distributed under the LFM Open License v1.0, which permits use, modification, and redistribution — including commercial use — for organizations with less than $10 million in annual revenue. Commercial use by larger companies is not covered by the license, requiring a separate arrangement with Liquid AI; qualified nonprofits are exempt from the threshold for non-commercial and research purposes.

Labonne framed the structure as a way to sustain model development — “the models are really the moats, so we need to be sensible in the way that we license them; otherwise, we cannot make money, so we can’t make more models” — while characterizing the threshold as a light-touch mechanism in practice.

Asked how the company would even know if a large enterprise quietly deployed the open weights, he was candid: “I think this is a question for our legal team, but personally, I don’t know. And even if you’re above $10 million, the only thing that we ask you is to contact us.”

The company pairs its licensed model releases with freely published research, he added, including new structured-output evaluations and a training technique that mitigates the repetition loops common in small models — a failure mode he noted Qwen models are “kind of guilty of.”

Small model, big enterprise implications

The launch coincided with an announcement from MacPaw, the Ukrainian software company behind CleanMyMac and Setapp, of a long-term strategic partnership with Liquid AI to build an on-device AI stack for the Mac.

Liquid AI will design and fine-tune foundation models for Eney, MacPaw’s macOS assistant, running locally on Apple silicon through MacPaw’s Elix inference engine and Mnemos memory layer, with results expected later this year.

Labonne pointed to the deal as a concrete validation of the size argument: “One of the reasons why they chose us is also because the model is quite small, and they don’t have all the memory budget to run the other models.”

The release arrives as hardware vendors, operating system developers, and enterprise software companies increasingly invest in local AI execution — and as agent harnesses proliferate across the industry. Liquid AI’s bet is that deployment economics, not raw scale, will define an important segment of that market: agents running continuously, everywhere, at zero marginal token cost.

Whether small, highly optimized agent models become a significant segment of enterprise AI will ultimately depend less on benchmark scores than on operational reliability. But Liquid AI’s latest release suggests the next competitive frontier is no longer simply building larger models — it’s building models small enough, and capable enough, to run wherever enterprise workflows already live.

More details continue to trickle out about OpenAI’s mysterious new hardware device — described previously as an AI-fueled smart speaker that will be the “physical manifestation” of ChatGPT.

Bloomberg now reports that the device will be “donut-shaped,” designed thusly to allow users to carry it around their home and place it in different locations, like a bedside table or a kitchen counter.

It will be constructed from “high-quality metal,” have a “premium look,” and (in a detail that mystifies) will have distinct “moving parts,” sources told Bloomberg.

It also could be slightly more expensive than your average smart speaker, perhaps $300 to $400 per unit, according to this report. For comparison, most of Amazon’s smart home speakers range in price from $40 on the low end to $240 on the high end.

So, to sum up: an expensive talking AI donut that has … moving parts? OpenAI releasing a smart home device has a certain logic to it, in that it would further integrate ChatGPT into users’ lives. However, historically speaking, smart speakers have not always been profitable and may prove a difficult market to break into. The potentially high price point also might not help.

The device, which is being developed in partnership with LoveFrom, the design studio founded by famous former Apple developer Jony Ive, will likely be released at some point in 2027, Bloomberg writes.

The company’s attempt to enter the hardware market has not gone off without a hitch. OpenAI is being sued by the current king of hardware, Apple, which has accused the AI lab of stealing trade secrets. OpenAI has denied wrongdoing.