Every co-founder Elon Musk recruited to build xAI has now reportedly left the company. Manuel Kroiss, who led the pretraining team, told people this month that he was departing. Ross Nordeen, described by Business Insider as Musk’s “right-hand operator,” left on Friday. They were the last two of eleven co-founders, all of whom have exited a company that was valued at $250 billion when SpaceX acquired it in February and that Musk himself described two weeks ago as having been “not built right the first time around.”
The departures are not ordinary startup attrition. The researchers Musk assembled in 2023 were among the most accomplished in artificial intelligence. Jimmy Ba co-authored the 2014 Adam optimisation paper, the most-cited paper in AI with more than 95,000 citations. Igor Babuschkin, the chief engineer, came from Google DeepMind. Christian Szegedy came from Google. Tony Wu led the reasoning team. Greg Yang, Toby Pohlen, Zihang Dai, Guodong Zhang, and Kyle Kosic brought experience from DeepMind, Google, Microsoft, and OpenAI. That entire cohort is now gone, and the company they helped build is being, in Musk’s words, “rebuilt from the foundations up.”
The exodus accelerated sharply in early 2026. Christian Szegedy left in February 2025, an early signal. But the cascade began in earnest when Tony Wu, one of the most operationally central co-founders, announced his departure on February 10, 2026. Jimmy Ba resigned within 24 hours, reportedly amid tensions over demands to improve model performance. By mid-March, only Kroiss and Nordeen remained. Their departures this week complete the sweep.
The timing is difficult to separate from the corporate restructuring happening around xAI. On February 2, SpaceX acquired xAI in an all-stock transaction that valued SpaceX at $1 trillion and xAI at $250 billion, creating a combined entity worth $1.25 trillion, the largest corporate merger by valuation in history. The deal brought xAI, X (formerly Twitter), and SpaceX under a single corporate umbrella, with SpaceX now preparing for a potential IPO in mid-2026 that could target a $1.75 trillion valuation.
The 💜 of EU tech
The latest rumblings from the EU tech scene, a story from our wise ol’ founder Boris, and some questionable AI art. It’s free, every week, in your inbox. Sign up now!
Weeks earlier, in January, Tesla invested $2 billion in xAI’s Series E round at an approximate $230 billion valuation. Tesla shareholders are suing Musk for breach of fiduciary duty over the investment, arguing that the company’s chief executive effectively directed shareholder capital into his own private venture. The lawsuit gained additional force on March 13, when Musk publicly acknowledged that xAI’s products, particularly its coding tools, were not competitive with Anthropic’s Claude Code or OpenAI’s Codex. Tesla had invested $2 billion in a company whose founder admitted it needed to be rebuilt from scratch.
Musk’s admission on March 13 was unusually candid for a chief executive whose company had just been acquired for a quarter of a trillion dollars. He said xAI’s AI coding tools simply did not work, and that the underlying system needed to be rebuilt. The statement appeared to validate the co-founders’ decision to leave: if the company’s own leadership acknowledges that the product failed, the researchers who built it have limited incentive to stay for the rebuild, particularly when they can command extraordinary compensation at competitors.
The AI talent market in 2026 is the most competitive it has ever been. Meta has reportedly offered packages worth up to $300 million over four years to retain top AI researchers. OpenAI, Google DeepMind, and Anthropic are all expanding their research teams aggressively. The eleven researchers who left xAI represent a concentration of talent that any of those companies would pay handsomely to acquire. Where they end up will say as much about the industry’s future direction as their departure says about xAI’s past.
xAI is not without assets. The Colossus supercomputer, built with more than 200,000 NVIDIA H100 GPUs, remains one of the largest AI training clusters in the world. Grok, the company’s chatbot, has a distribution channel through X’s user base. And the SpaceX merger provides access to capital, infrastructure, and engineering talent at a scale that few AI companies can match. The question is whether infrastructure and distribution are sufficient when the research leadership that was supposed to make the product competitive has entirely departed.
The xAI co-founder exodus follows a pattern that has repeated across Musk’s companies. Twitter lost the majority of its senior leadership and roughly 80 per cent of its workforce within months of his 2022 acquisition. Tesla’s senior ranks have thinned steadily as Musk’s attention has divided across six companies. The common thread is a management style that produces extraordinary results in hardware engineering, where Musk’s tolerance for risk and pace of iteration have built SpaceX and Tesla into industry-defining companies, but appears less effective in research-driven fields where the most valuable people have abundant alternatives and low tolerance for instability.
Artificial intelligence research is, in 2026, the most competitive labour market in technology. The researchers who co-founded xAI did not need to be there. They chose to be, attracted by the resources Musk could deploy and the ambition of the project. That every one of them has now chosen to leave, during a period when the company received a $250 billion valuation and access to the resources of SpaceX, suggests that the problems at xAI are not principally financial or infrastructural. They are organisational. And no amount of capital can rebuild a research culture once the people who created it have gone.
Apple Card users who subscribe to Uber One can now earn up to $30 in Uber Cash on eligible Uber Eats grocery and retail purchases. Here’s how to get it.
The promotion provides up to $30 in Uber Cash through August 31. Uber One members can earn $10 on one qualifying order each month in June, July, and August when they pay with Apple Card in the Uber Eats app.
Eligible purchases include groceries, health and beauty products, pet supplies, and other retail items available through Uber Eats. Uber Cash earned through the promotion can be used on future Uber Eats orders.
The promotion expands the benefits available to Apple Card users who regularly use Uber’s services. Customers who already subscribe to Uber One and place grocery or retail orders through Uber Eats can combine the monthly Uber Cash bonus with the card’s existing rewards.
The offer is separate from Apple Card’s standard rewards program. Customers who use Apple Card with Apple Pay for Uber and Uber Eats purchases continue to receive 3% Daily Cash on eligible transactions.
Uber and Uber Eats are among the temporary merchants that offer 3% Daily Cash on eligible Apple Pay purchases made with Apple Card. Participating, long-term partners, include Nike, Ace Hardware, Hertz, Walgreens, and Exxon Mobil.
Uber One is the company’s subscription service that includes benefits such as discounts on eligible orders and reduced delivery fees. Apple Card customers are also eligible for a six-month Uber One trial when signing up with Apple Card and Apple Pay.
After the trial period ends, the subscription automatically renews at Uber’s standard monthly rate unless canceled.
Atlas Menu, a GTA V cheat service that promised “enhanced privacy,” was hacked, exposing nearly 64,000 accounts including emails, usernames, hashed passwords, and IP addresses. The stolen data was posted to GitHub by a hacker motivated by revenge.
TL;DR
Atlas Menu, a cheat service for Grand Theft Auto V’s online mode, has been hacked, exposing the personal data of nearly 64,000 users. The stolen data included email addresses, usernames, hashed passwords, IP addresses, and support tickets, according to data breach notification service Have I Been Pwned. The breach was claimed by a hacker whose stated motivation was revenge against a scammer, and the allegedly stolen data was posted publicly to GitHub.
The irony is difficult to miss. Atlas Menu marketed itself on security, promising users “secure authentication and enhanced privacy through our advanced encryption techniques,” according to its official website, which was archived before going offline. The service sold features that gave GTA V players unfair advantages: invisibility, super jumps, the ability to fly through the map, and other modifications to Rockstar Games’ online multiplayer environment.
Game cheats have evolved from hobbyist projects into a multi-million dollar commercial market. Professional and semi-professional gamers pay for software that provides advantages over competitors, while casual players use cheats to bypass progression systems or grief other users. Services like Atlas Menu operated in a legal grey area, selling tools that violate most games’ terms of service but are not themselves illegal in most jurisdictions.
The breach exposes users to potential embarrassment and further security risks. Email addresses tied to gaming accounts often overlap with primary personal accounts, meaning that users who signed up for a cheat service with the same credentials they use elsewhere now face credential-stuffing attacks. The hashed passwords provide some protection, but depending on the hashing algorithm used, determined attackers could crack weaker passwords.
Atlas Menu is not the first cheat service to be breached. A popular cheat service for Counter-Strike: Global Offensive was hacked several years ago in a similar incident. The GTA franchise has been a recurring target for security incidents, from the 2023 GTA VI trailer leak traced to a Rockstar developer’s son to repeated exploits targeting GTA Online’s peer-to-peer networking architecture.
The breach highlights a structural problem in gaming security. Cheat services operate with minimal accountability: their operators are often anonymous, their infrastructure is disposable, and their users have no recourse when things go wrong. Unlike enterprise software vendors that face regulatory and contractual obligations to protect customer data, cheat service operators answer to no one.
Rockstar Games has invested heavily in anti-cheat systems for GTA Online, but the persistence of services like Atlas Menu demonstrates the limits of technical enforcement against a motivated and commercially incentivised cheating ecosystem. Law enforcement has targeted some gaming-adjacent cybercrime operations, but cheat services rarely rise to the level of criminal prosecution unless they involve direct financial fraud or intellectual property theft at scale.
The owners of Atlas Menu could not be reached for comment. The site remains offline.
For years, a unit of Russia’s military intelligence agency quietly turned ordinary home routers into tools of espionage. The GRU group known as APT28, the same outfit behind the 2016 DNC hack and a string of attacks on NATO targets, exploited unpatched firmware and unchanged default passwords to compromise thousands of devices across 23 US states, redirecting internet traffic through servers under Russian control and harvesting credentials along the way. Federal agents disrupted the operation in April under a court order. What they couldn’t do from a distance was fix the underlying vulnerabilities. That requires five steps from you.
The attack targeted small-office/home-office routers, also known as SOHO routers, and was carried out by a unit in the Russian military intelligence agency, the GRU. Government agencies are urging people to follow basic router hygiene steps, such as updating to the latest firmware and changing default login credentials. The UK’s National Cyber Security Centre includes a number of TP-Link routers specifically targeted by the hackers.
While that news sounds pretty alarming, it’s worth keeping in mind that the attack compromised enterprise routers specifically, so your home Wi-Fi router likely isn’t at risk. That said, some of the affected routers can be used as standard home routers, so it’s worth checking whether your model was exploited in the attack.
“There is a big trend of exploiting routers these days, and that goes both for the consumer and enterprise or corporate routers,” Daniel Dos Santos, vice president of research at the cybersecurity company Forescout, told CNET.
A news release from the NSA notes that the attack indiscriminately targeted a wide pool of routers, with the goal of gathering information on “military, government, and critical infrastructure.”
This attack is linked to threat actors within the Russian GRU — which go by APT28, Fancy Bear, Forest Blizzard and other names — and has been ongoing since at least 2024, according to the FBI.
It’s known as a Domain Name System hijacking operation, in which DNS requests are intercepted by changing the default network configurations on SOHO routers, allowing the actors to see a user’s traffic unencrypted.
“For nation-state actors like Forest Blizzard, DNS hijacking enables persistent, passive visibility and reconnaissance at scale,” says a Microsoft Threat Intelligence report on the attack.
Microsoft identified more than 200 organizations and 5,000 consumer devices impacted by the GRU’s attack.
The FBI’s announcement refers to one router specifically, the TP-Link TL-WR841N, a Wi-Fi 4 model that was originally released in 2007. The UK’s National Cyber Security Centre lists 23 TP-Link models that were targeted, but notes that it is likely not exhaustive.
Here is the list of affected devices:
A TP-Link Systems spokesperson told CNET in a statement that the affected models all reached End of Service and Life status several years ago.
“While these products are outside our standard maintenance lifecycle, TP‑Link has developed security updates for select legacy models where technically feasible,” the spokesperson said.
TP-Link is urging people with these outdated routers to upgrade to a newer device if possible. You can find a list of available security patches on its security advisory page addressing the recent attack.
The NSA referred organizations to a list of best practices for securing your home network. The most important thing you can do if you’re using one of the impacted devices is to upgrade your router as soon as possible. It likely hasn’t received firmware updates in years, which is like leaving the door to your network unlocked.
“The longer you carry on doing that, the greater the risk,” said Rik Ferguson, vice president of security intelligence at Forescout. “The router sits in such a privileged position within any network. All of your communication, all of your traffic, has to pass through that device.”
In addition to using a newer device that’s still getting security updates, there are a few other steps you can take to lock down your network:
‘The End of Oak Street’ opens with a quiet drive down a familiar street that turns into something no one could have prepared for. The Platt family heads out as usual until the road simply stops. Instead of continuing into town, it ends at a cliff that drops away into an entirely different landscape. Their neighborhood has moved, and nothing around them matches the world they left behind.
Confusion spreads like wildfire among the people, as Denise Platt, played by Anne Hathaway, tries to make sense of what’s going on for her son, who is asking the question that everyone is asking: what in the world is going on. She informs him that their house, street, and almost everything else has moved overnight. The family quickly realises they are not the only ones in this strange new environment. A deafening roar cuts through the air, and the first massive dinosaur appears where lawns and mailboxes used to be the edge of safety.
Sale
The Platts and their neighbors are then put on a merry-go-round that is constantly in motion. More and more dinosaurs start to arrive & suddenly the transplanted suburban blocks turn into an open hunting ground. Backyards, driveways, and living rooms, where the youngsters used to play, have become part of the danger zone. Denise and Greg, played by Ewan McGregor, make every decision with one thing in mind: keeping their two children alive. Christian Convery and Maisy Stella bring the children to life, and the fact that they are always in danger makes each close encounter much more horrifying.
David Robert Mitchell wrote and directed the film. He reunited with several frequent collaborators, including cinematographer Michael Gioulakis, production designer Maya Shimoguchi, editor John Axelrad, and composer Michael Giacchino. Their work together gives the story a consistent feel that blends everyday neighborhood details with the scale of prehistoric creatures moving through them.
J.J. Abrams is producing through his Bad Robot company, alongside Hannah Minghella and a few others. Warner Bros. Pictures are set to release it in theaters and IMAX on August 14. This one screams J.J. Abrams, with its huge, mysterious, event-scaled scope, yet it’s also firmly grounded in one family’s desperate effort to make it through each day.
Canada’s Prime Minister, Mark Carney, announced in May 2026 that the nation’s next-gen spy planes will no longer come from the United States. Instead of purchasing aircraft from Boeing, the nation has decided to buy radar planes from Swedish company Saab (yes, the Saab that used to make cars).
As reported by outlets such as WRAL News, Carney stated that the nation had entered negotiations with Saab to purchase its Airborne Early Warning & Control Aircraft, the GlobalEye, which is based on the Canadian-made Bombardier Global 6500. In the May 24 announcement, Carney said that “Saab’s GlobalEye will be a key resource for the Canadian Armed Forces to detect and deter threats across the Arctic.”
GlobalEye is essentially a flying radar system, capable of detecting threats in the air, on land, and at sea. It can fly for more than 12 hours and has a radar range of over 400 miles. Its Erieye ER radar can detect small targets, even in very cluttered conditions, while its Ground Moving Target Indication can identify moving objects over large distances. Programs like the Automatic Identification System can then, in turn, identify those spotted objects.
Canada has purchased spy planes from the U.S. before, acquiring a fleet of Boeing aircraft in 2023 to replace its then-aging fleet. It even considered Boeing’s E-7 Wedgetail this time around. The nation’s decision to go with Saab and Bombardier is part of an attempt to reduce dependence on the United States. In March 2026, Prime Minister Mark Carney stated that Canada intended to take on the responsibility of protecting its Arctic territory on its own, moving away from the decades-long partnership with the U.S. it had previously relied on.
This move comes amid the backdrop of increased tensions between the United States and Canada, not least due to President Donald Trump’s threat to impose 100% tariffs on Canadian imports and his decision to revoke Carney’s invitation to join the Peace Council. Canada’s decision to purchase the GlobalEye is expected to help create jobs in Canada, as the aircraft is based on a Canadian Bombardier plane and uses the same supply chain. Saab’s reconnaissance aircraft have been used by other nations, too, with Sweden having sent Saab ASC 890 planes to Ukraine in 2024.
Anthropic has confidentially filed S-1 paperwork for an IPO, potentially beating rival OpenAI to the public market this fall. The filing follows a $65 billion funding round at a $965 billion valuation and projected Q2 revenue of $10.9 billion, with Anthropic on pace for its first profitable quarter.
TL;DR
Anthropic has confidentially submitted draft registration paperwork for an initial public offering with the Securities and Exchange Commission, the company announced in a blog post on Monday. The number of shares to be offered and the price have not been set. The filing positions Anthropic to potentially reach the public market as soon as this fall, ahead of rival OpenAI, which is also preparing its own confidential IPO filing in the coming weeks.
The move comes days after Anthropic raised $65 billion in a funding round at a $965 billion valuation, eclipsing OpenAI’s valuation for the first time. The round was led by Altimeter Capital, Dragoneer, Greenoaks, and Sequoia Capital. Goldman Sachs, JPMorgan Chase, and Morgan Stanley are expected to be under consideration for lead underwriting roles on both Anthropic’s and OpenAI’s listings.
Anthropic’s financial growth over the past 12 months has been extraordinary even by AI startup standards. The company’s annualised revenue run rate was $4 billion in July 2025. By January 2026 it had surpassed $9 billion. Anthropic has told investors that the run rate will exceed $50 billion by the end of July, representing roughly an 80-fold increase in annualised revenue over two years.
The company expects to report $10.9 billion in revenue for the second quarter of 2026, more than doubling the $4.8 billion it generated in Q1 and exceeding its entire 2025 annual revenue in a single quarter. Anthropic is also on pace for its first profitable quarter, with a projected operating profit of $559 million. Early investors like Accel, whose Anthropic stake has more than quadrupled in months, now have a clear path to liquidity.
The 💜 of EU tech
The latest rumblings from the EU tech scene, a story from our wise ol’ founder Boris, and some questionable AI art. It’s free, every week, in your inbox. Sign up now!
The growth has been driven by a surge in enterprise demand for Claude’s coding capabilities and, more recently, by Anthropic’s Mythos cybersecurity model, which discovered more than 10,000 zero-day vulnerabilities across major operating systems and has become the most discussed AI capability in enterprise security.
Anthropic’s filing sets up what could be the most significant cluster of technology IPOs since the dot-com era. OpenAI is preparing its own confidential filing and has targeted a public debut in the fall at a potential valuation of up to $1 trillion. SpaceX has already filed for the largest IPO in history, with marketing set for early June. All three companies are backed by overlapping pools of sovereign wealth funds and institutional investors, and all three are likely to draw on Goldman Sachs, JPMorgan, and Morgan Stanley for underwriting.
For Anthropic, filing first carries a strategic advantage. The company that reaches the public market ahead of its rival gets first access to the broadest possible pool of institutional and retail investors, sets the valuation benchmark for the AI sector, and defines the narrative around which AI company represents the best public market bet.
The competitive dynamic between the two companies has intensified this year. While Anthropic has posted an 80-fold annualised revenue increase and surpassed OpenAI in valuation, OpenAI has been rethinking its product lineup, shuffling leadership, and confronting reports that it missed certain internal revenue and user targets. OpenAI CFO Sarah Friar has pushed back, telling Bloomberg the company is seeing a “vertical wall of demand” for its products.
Anthropic’s S-1, when it becomes public, will need to disclose several material risks. The most prominent is the company’s ongoing legal battle with the US government after the Pentagon declared Anthropic a supply-chain risk, a designation typically reserved for foreign adversaries. The dispute stems from Anthropic’s refusal to grant the military unrestricted access to its models. Anthropic has said the designation could jeopardise billions of dollars in revenue.
The $965 billion valuation also carries scrutiny. At $10.9 billion in projected Q2 revenue, Anthropic would be trading at roughly 22x annualised revenue, a premium that assumes continued hypergrowth. If the $50 billion run rate materialises by July, the multiple compresses to a more defensible 19x, but that figure depends on maintaining the pace of enterprise adoption that drove the Q1-to-Q2 doubling.
There is also the question of whether profitability is sustainable. Anthropic’s projected $559 million operating profit in Q2 represents a roughly 5% margin, thin for a company seeking a near-trillion-dollar public valuation. The economics of running frontier AI models at scale remain expensive, and compute costs could compress margins as the company scales. The University of Michigan turned a $20 million early OpenAI investment into $2 billion, a reminder that early AI bets can produce extraordinary returns, but public market investors will demand more predictable economics.
Anthropic’s IPO filing, combined with OpenAI’s expected filing and SpaceX’s already-submitted paperwork, means the fall 2026 IPO window could see more than $200 billion in new public market value from three companies alone. The last time a cluster of technology offerings this large reached the public market simultaneously was the 1999-2000 dot-com wave.
The comparison is imperfect. Anthropic and OpenAI have real, rapidly growing revenue. SpaceX is profitable. But the valuations all assume that the AI market will continue expanding at its current pace, that enterprise adoption will not plateau, and that regulatory and legal risks will remain manageable. Anthropic’s confidential filing is the starting gun for the most consequential test of whether the AI boom can sustain public market scrutiny.
Big news in enterprise AI broke over the weekend as Chinese AI startup MiniMax released its highly anticipated M3 large language model on Sunday evening Eastern time, pairing frontier-tier coding and agentic performance with a 1-million-token context window and native multimodality for a fraction of the cost of leading proprietary models, with pricing starting at just $20 per month under its new subscription token plans.
The company’s leadership also announced plans to deliver the model under an open source license including “open weights,” allowing for full enterprise downloading and customizability free-of-charge, coming sometime in the next 10 days. For now, it is available via the MiniMax API at a special discounted price of $0.3 per 1 million input tokens and $1.20 per million output tokens (on fresh cache) for the next week — beating proprietary U.S. giants like Google, OpenAI and Anthropic handily on cost, while also eclipsing the performance of the latest models from the former two on selected benchmarks.
Even at its full price of $0.6/$2.40 per million input/output tokens, MiniMax-M3 remains at just 8-20% the cost of the leading, proprietary U.S. models.
The traditional matrix governing large language model development has long dictated a rigid choice: software developers can either access top-tier closed-source intelligence behind restrictive APIs, or deploy nimble, cost-effective open models that falter on multi-step reasoning, dense coding tasks, and massive data sequences. MiniMax-M3 fundamentally upends this paradigm.
By unifying these two historically separated frontier capabilities, M3 introduces a level of comprehensive utility previously restricted to expensive, closed-source ecosystems, effectively shifting the baseline of open-weights systems while drastically minimizing the operational compute footprint required to execute complex development loops.
|
Model |
Input |
Output |
Total Cost |
Source |
|
MiMo-V2.5 Flash |
$0.10 |
$0.30 |
$0.40 |
|
|
deepseek-v4-flash |
$0.14 |
$0.28 |
$0.42 |
|
|
deepseek-v4-pro |
$0.435 |
$0.87 |
$1.305 |
|
|
MiniMax-M3 |
$0.30 |
$1.20 |
$1.50 (limited time only) |
|
|
Gemini 3.1 Flash-Lite |
$0.25 |
$1.50 |
$1.75 |
|
|
MiMo-V2.5 |
$0.40 |
$2.00 |
$2.40 |
|
|
Grok 4.3 low context |
$1.25 |
$2.50 |
$3.75 |
|
|
GLM-5 |
$1.00 |
$3.20 |
$4.20 |
|
|
Kimi-K2.6 |
$0.95 |
$4.00 |
$4.95 |
|
|
GLM-5.1 |
$1.40 |
$4.40 |
$5.80 |
|
|
Grok 4.3 high context |
$2.50 |
$5.00 |
$7.50 |
|
|
Qwen3.7-Max |
$2.50 |
$7.50 |
$10.00 |
|
|
Gemini 3.5 Flash |
$1.50 |
$9.00 |
$10.50 |
|
|
Gemini 3.1 Pro Preview ≤200K |
$2.00 |
$12.00 |
$14.00 |
|
|
GPT-5.4 |
$2.50 |
$15.00 |
$17.50 |
|
|
Gemini 3.1 Pro Preview >200K |
$4.00 |
$18.00 |
$22.00 |
|
|
Claude Opus 4.8 |
$5.00 |
$25.00 |
$30.00 |
|
|
GPT-5.5 |
$5.00 |
$30.00 |
$35.00 |
At the core of the model’s efficiency lies an architectural departure from classic Transformer networks. Standard attention mechanisms scale quadratically ($O(N^2)$), meaning computational and financial costs explode as text inputs lengthen.
To combat this “inherent flaw,” the engineering team implements MiniMax Sparse Attention (MSA), a clean, extensible sparse attention blueprint.
To visualize this innovation, think of traditional full attention as an editor reading an entire library from scratch every time they need to verify a single sentence. MSA acts as an intelligent indexing clerk, using a pre-filtering phase to partition Key-Value (KV) matrices into highly precise blocks.
At the operator level, MSA uses a “KV outer gather Q” approach. The system treats KV blocks as an outer loop, dynamically aggregating only the specific queries that hit them. Because each data block is read exactly once and memory access remains strictly contiguous, hardware utilization skyrockets.
In internal trials, MSA runs more than 4x faster than alternative open-source solutions like Flash-Sparse-Attention or flash-moba.
When managing a maxed-out context length of 1 million tokens, M3’s per-token compute demand drops to just 1/20th of the previous generation model, translating into a 9x acceleration in the prefilling stage and a 15x boost during decoding.
Rather than taking a pretrained text network and fusing it with a separate vision model, MiniMax engineered M3 as a natively multimodal system from “Step Zero”.
The company overhauled its data ingest machinery to blend naturally interleaved sequences of text, images, and visual components, scaling the total pretraining corpus beyond 100 trillion tokens.
This deep data alignment enables the model to translate complex visual geometries, such as programming charts or coordinate maps, into structural code without losing contextual fidelity. On standardized assessments, M3 validates this engineering path.
The model records a 59.0% on SWE-Bench Pro, an autonomous agent metric, positioning it ahead of closed models like GPT-5.5 and Gemini 3.1 Pro. It achieves a 66.0% on Terminal Bench 2.1, a 74.2% on MCP Atlas, and an 83.5 on BrowseComp—outstripping Claude Opus 4.7’s benchmark score of 79.3 in autonomous browsing and information retrieval.
However, when contrasted with Anthropic’s newly released, premium frontier model, Claude Opus 4.8, from last week, the competitive ceiling of M3’s efficient sparse-attention footprint becomes evident across directly comparable, tool-intensive agent benchmarks.
In the domain of pure code modification on SWE-Bench Pro, M3’s 59.0% score drops behind Opus 4.8’s leading 69.2% threshold.
A similar performance delta manifests in automated system environments via Terminal-Bench 2.1; while M3’s 66.0% terminal execution score effectively runs neck-and-neck with the previous-generation Opus 4.7 baseline of 66.1%, it trails the upgraded Opus 4.8 architecture, which achieves 74.6%.
Furthermore, evaluations tracking continuous GUI interaction on the OSWorld-Verified sandbox place M3’s automated computer use at 70.0%, compared to a higher 83.4% validation rate secured by Opus 4.8.
These standardized evaluations illustrate the structural trade-offs currently defining the ecosystem: closed-source systems like Opus 4.8 maintain absolute margin leads on hyper-complex reasoning vectors, yet M3 delivers a highly capable baseline of local, tier-one automated operation without the compounding premium of closed-door API subscription fees.
When positioned alongside the heavy-duty inference metrics of the newly minted, fellow open weights model DeepSeek-V4 Pro Max, M3 holds its ground across core agentic categories while asserting narrow advantages in specialized code synthesis.
On the software engineering matrix of SWE-Bench Pro, M3’s 59.0% resolution efficiency edges past DeepSeek-V4 Pro Max’s score of 55.4%.
However, the competitive friction tightens in command-line environments; under Terminal Bench evaluations, DeepSeek-V4 Pro Max pulls slightly ahead with a 67.9% execution accuracy over M3’s 66.0% mark.
In web orchestration and open-world browsing simulations, the two architectures reach a virtual statistical parity, with M3 registering an 83.5% on BrowseComp compared to DeepSeek’s 83.4%.
Similarly, on the MCP Atlas tool-use framework, M3 secures a narrow lead at 74.2% against DeepSeek’s 73.6%.
This close alignment demonstrates that while DeepSeek handles a massive 1.6-trillion total parameter footprint with specialized high-effort reasoning modes, MiniMax’s block-filtered sparse attention mechanism yields directly competitive execution efficiencies without requiring extensive parameter activation scaling.
MiniMax translates these architectural gains into immediate utility through an updated product suite divided between standalone applications, customizable subscription tiers, and raw developer infrastructure. For end-user orchestration, the flagship implementation is MiniMax Code, an AI agent product designed to maximize M3’s multi-step capabilities.
Operating via web or native desktop apps, MiniMax Code runs an “Agent Team” capable of breaking massive engineering tasks into multi-stage, concurrent workflows.
The system relies on a “Producer + Verifier” adversarial harness loop. As one agent instance generates code, a secondary verifier instance aggressively tests and reflects upon execution outputs, allowing the network to self-correct and operate autonomously for days without human oversight. Because of its native visual grounding, MiniMax Code supports direct computer use.
A developer can issue a cross-application voice prompt via their phone to have the model open a localized enterprise ERP client and batch-populate data tables directly from an open Excel spreadsheet.
For custom setups, developers can pipeline M3 directly into existing workflows using an API key (sk-cp) compatible with common alternative IDE environments like Claude Code, Cursor, Roo Code, and Cline. The API introduces a toggleable “thinking mode”.
When enabled, M3 routes processing power into deep reasoning and long-horizon planning; when disabled, the model runs at minimal latency for quick text completion. The companion Token Plan models an aggressive pricing strategy structured around shared multimodal quotas. Billed annually, three options are available:
Plus ($20/month): Supplies ~1.7B tokens per month and handles 3–4 concurrent agents.
Max ($50/month): Supplies ~5.1B tokens per month, manages 4–5 concurrent agents, and adds 3 automated video clips per day via Hailuo 2.3.
Ultra ($120/month): Supplies ~9.8B tokens per month, facilitates 6–7 concurrent agents, and extends video capacity to 5 daily clips.
MinMax’s pledge to release M3 under an open-weights license model—with weights and technical documentation launching on HuggingFace and GitHub within 10 days—carries significant strategic weight for enterprise infrastructure managers.
However, it is still to be determined precisely which license the weights will be available under, and whether or not it will be permissible for consumer usage, e.g. MIT, Apache 2.0 or the new OpenMDW license. If so, the calculus looks like this:
|
Feature / Model Attribute |
Closed API Providers (e.g., GPT-5.5, Opus 4.7) |
Open-Weights Frontier (MiniMax M3) |
|
Data Privacy & Boundaries |
Requires external API requests; potential data ingestion vectors. |
Total local isolation; runs entirely inside private user clusters. |
|
Custom Optimization |
Limited to basic fine-tuning wrappers or prompt engineering. |
Full pipeline control; architecture allows deep adapter/weights customization. |
|
Cost Vector Consistency |
Bound to perpetual per-token API pricing models. |
Computational demands cut to 1/20th; mitigates hardware ceiling. |
By shipping the underlying model weights directly to the community, MiniMax departs from the closed-door approach favored by major American AI labs.
For enterprise users bound by strict compliance and privacy rules, open weights mean they can run M3 locally on internal hardware.
This setup completely removes the risk of data leakage associated with public APIs. Furthermore, it permits engineering teams to run bespoke fine-tuning passes, modify internal architectures, or embed specialized system prompts deep within the model layers—transforming an off-the-shelf system into a highly targeted proprietary asset.
The developer ecosystem reacted immediately to M3’s operational benchmarks, singling out its long-horizon autonomous behavior and cost-to-performance profile.
A major focal point of discussion is a 12-hour automated verification test where M3 was tasked with reproducing an ICLR 2025 Outstanding Paper Award winner, titled “Learning Dynamics of LLM Finetuning”.
As MiniMax’s own researcher @MikaStars39 highlighted on X:
“M3 ran autonomously for nearly 12 hours, producing 18 commits and 23 experimental figures on its own, and got the core experiments working:
it matched the predicted probability trends in the SFT stage
clearly observed the squeezing effect central to the DPO experiments
validated the Extend mitigation method proposed in the original paper.”
Simultaneously, creators of developer tools highlighted the practical economic advantages of the model’s new attention mechanism. The official team behind the agentic AI coding harness Cline posted an alert confirming day-one compatibility, stating:
“The new MiniMax-M3 is their first model to have 1m context, multimodal, and agentic coding capability. Congratulations to @MiniMax_AI for the breakthrough in sparse-attention architecture cutting compute & cost to 1/20th their previous generation.”
This sharp drop in execution costs shifts how developers view the relationship between financial investment and capability. Tech commentator @jumperz mapped out this disruption, noting how M3 breaks a historical pattern in machine learning pricing:
By addressing context scaling limitations through fundamental attention-level optimizations rather than brute-force hardware scaling, MiniMax has established a highly efficient open-source baseline. M3 demonstrates that the next phase of agent development will not just be driven by larger datasets, but by efficient architectural choices that make frontier-level performance accessible to the broader open-source community.
For enterprises building autonomous software development or agent infrastructure, MiniMax M3 provides the ultimate “bang for the buck.”
While DeepSeek-V4 Pro holds a microscopic price advantage of $0.195 per million tokens, MiniMax M3 justifies its marginal premium by delivering superior autonomous software engineering resolution rates (59.0% SWE-Bench Pro).
More importantly, because M3 is an open-weights model, the calculation extends far beyond the API chart. By deploying M3’s weights locally inside private enterprise clouds, organizations completely bypass cloud data egress tracking, eliminate structural vendor lock-in, and can implement custom prefix-caching models on internal hardware. This technical approach transforms a highly efficient runtime budget into a permanent, privately owned corporate asset.
Expect more than an hour of trailers and reveals.
Sony is kicking off the not-E3/Summer Game Fest season with a State of Play event on June 2 at 5PM ET. You can tune in live via the company’s official YouTube and Twitch channels. A version with Japanese subtitles will be available on YouTube.
We are parking the YouTube stream below. You can keep the page open and press play when the stream starts.
The company has promised more than an hour of footage, with “updates, announcements and gameplay reveals from top studios around the world.” We don’t know exactly what will be revealed, but we do know that the event will highlight Insomniac’s highly-anticipated Marvel’s Wolverine game. This makes sense, as the PS5 exclusive will be available to play in just over two months on September 15.
We don’t know much beyond everyone’s favorite long-clawed mutant, so let’s head into speculation territory. Rumors have suggested that the stream may give us some new details on Naughty Dog’s upcoming title, Intergalactic: The Heretic Prophet. There’s also been rumors of some kind of God of War spinoff coming down the pike. The upcoming PS5 games Phantom Blade Zero and Marvel Tokon: Fighting Souls are both coming out soon, so will likely get final trailers.
That leaves plenty of room for first-party surprises, which is something Sony very much needs right now. PS5 sales are down nearly 50 percent year-over-year, which should come as no surprise. Console prices keep going up and people’s personal finances keep going down. Sony’s latest console is six years old and is now more expensive than it was at launch. In the past, gamers could expect steep discounts by the six-year mark, which usually led to an uptick in sales.
Given current geopolitical concerns and AI’s never-ending hunger for memory, it’s unlikely Sony will lower prices anytime soon. That leaves one tried-and-true option to lure in new customers: make games people want to play.
Nearly 2,000 WordPress websites were infected with malware that relies on Steam Community profile comments to hide command-and-control (C2) data.
The threat actor used invisible Unicode characters to encode a payload that builds a URL to a malicious script. By leveraging Valve’s platform, the attacker avoids maintaining a separate C2 infrastructure and evades traditional detection methods.
Since the campaign was first uncovered in July 2025, GoDaddy security engineers have found malware on approximately 1,980 WordPress websites.
It is unclear how the hackers breach the websites, but researchers assess that the initial infection vector ranges from stolen admin logins or compromised FTP/SFTP credentials to the exploitation of a vulnerable WordPress theme or plugin, or a supply-chain compromise.
The first-stage malware planted on a website uses WordPress page loads to reach specific Steam profiles and extract text from benign-looking comments.
However, the text includes hidden Unicode characters that conceal malicious payloads sometimes disguised as ASCII art.
GoDaddy researchers note in a report that the threat actor uses six invisible Unicode characters for the encoded payload:
The decoder ignores any visible character and maps the invisible ones to a corresponding number; then it converts them to binary representation and reconstructs bytes from the binary stream.
“This encoding allows binary data to be embedded within normal-looking text. The visible characters serve as camouflage while the invisible characters carry the actual payload,” GoDaddy says.
According to the researchers, the decoded payload is used to build a hello-mywordl[.]info URL serving JavaScript code that is injected into every frontend WordPress page.
Based on the file names (e.g., asahi-jquery-min-bundle and lodash.core.min.js), the retrieved malware is disguised as a legitimate JavaScript library.
The final stage of the attack is implementing a backdoor that responds to specially crafted POST requests that include a specific authentication cookie. If the “tEcaKKXEsb cookie is present, the backdoor accepts base64-encoded PHP code via POST parameter,” the researchers explain.
GoDaddy describes several evasion mechanisms employed by the malware, including obfuscated strings using octal and hex escapes, randomized function names, fake disabled logging code, and the use of standard WordPress APIs, allowing it to blend with normal activity.
Site owners can defend by checking for references to Steam Community URLs, suspicious external JavaScript injections, outbound connections from WordPress servers to Steam, and unexpected scripts loading from domains such as hello-mywordl[.]info.
Other indicators include invisible Unicode characters, suspicious _transient_caption_ cache entries, disabled SSL verification in cURL requests, and POST requests containing the malware’s authentication cookies or the new_code parameter.
The researchers recommend that security teams prioritize restoring from a known good backup before the infection date. If this is not possible, the manual cleaning process should be thorough because “attackers can reinstall removed code through the backdoor if any component remains active.”
Automated pentesting tools deliver real value, but they were built to answer one question: can an attacker move through the network? They were not built to test whether your controls block threats, your detection rules fire, or your cloud configs hold.
This guide covers the 6 surfaces you actually need to validate.
CBP Commander-at-Large Gregory Bovino made that title literal by showing up wherever Trump needed trouble started. Once he had arrived far north of the southern border he was supposed to be patrolling, Bovino (and the people he was “commanding”) found themselves on the receiving end of several lawsuits.
Not only did they find themselves on the receiving end of lawsuits, they — especially Gregory Bovino — found themselves hit with judgments and orders forbidding them from constantly violating the rights of anti-ICE protesters and journalists covering the protests.
After a couple of murders were committed by CBP officers in Minneapolis, Minnesota, the Trump administration decided Bovino was more trouble than he was worth. Sure, he was loyal and loved to personally engage in violence against protesters, but he also loved to see himself on TV and to dress like he’s auditioning for a Hitler Youth leadership position.
Now that he’s back at the border and bored, Bovino appears to be using his free time to push his own personal brand: a buzzcut bigot willing to spread hatred wherever it’s welcomed. Jeff Tischauser points out on Bluesky that Bovino recently headed overseas to help European bigots push their anti-migrant narratives.
Greg Bovino will speak at a white nationalist conference in Portugal tomorrow. He will share the stage with no fewer than five people who idolize Hitler, including one who joined a group created by two Nazi SS members. Another guy is a self-described racist who refers to women as “cockroaches.‘
RESUM26 is this year’s “Remigration Summit,” which was held in Porto, Portugal on May 30. If you’re not familiar with “Remigration” and/or RESUM26, I’ll let the organization speak for itself, even if it can’t seem to limit itself to 14 words.
Remigration is the umbrella term that designates and encompasses a set of fiscal, cultural, economic, social, political, and logistical policies whose objective is to prevent population replacement through the reversal of migratory flows, thereby restoring the sovereignty, independence, and identity of countries, through the defense of their ethnocultural specificity.
It’s almost twice as long as it needs to be at 25 words, but it’s pretty much saying the same thing Nazi supporters have been saying for years. Oh, and since literal Nazi supporters spoke at this event, here’s another reminder that Bovino himself seems to be on the supply side of Nazi sympathy.
If you can’t see the Bluesky post, It’s a screencap of Bovino’s recent X post where he’s captured giving what looks a hell of a lot like a Nazi salute while in his CBP work uniform. The accompanying text suggests Nazi salutes are just another way federal officers can visibly show their support for ICE and its activities.
Here’s one of the multiple anti-migrant luminaries Bovino joined at RESUM26: Alfonso Goncalves, as summarized by the Global Project Against Hate and Extremism:
Gonçalves is a white supremacist and misogynist who fashions himself as an authoritarian leader. He revels in descriptions of himself as “transphobic,” and proudly accepts the characterization that he is “ultranationalist, racist, and xenophobic.” Gonçalves is particularly known for his misogyny, posting bizarre rants about women’s right to vote, casual sex, and women sitting in public. He refers to women as “whores” and “cockroaches.” He has argued that women who get divorced should “not be entitled to receive money/goods” from their husbands and should be mandated to “pay for damages caused to the family.” He says abortions are a “crime against humanity” and has called for women who have the procedure to receive the death penalty. For Gonçalves, “non-traditional families” are an “aberration,” “80% of all divorces are initiated by women” and people of African descent are creating a “population replacement” of people of “native” European descent. He has argued that “African American men are 12x more likely to commit murder than white men.” His bigotry is so extreme that even Elon Musk’s Twitter, known for being lenient towards hateful accounts, permanently suspended both his main and backup accounts, in one instance for “abusive behavior.”
And that’s just one speaker at this event. Also speaking at RESUM26 were Dutch far-right activist Eva Vlaardingerbroek (great replacement theory proponent), Austrian far-right activist Martin Sellner (great replacement theory proponent), former French National Front politician Jean-Yves Gallou (more of the same), and RESUM co-founder Dries Van Langenhover (who adds some Holocaust denialism to the mix).
Greg Bovino has done this while still employed by the US federal government. Under any normal president, his resignation might have been demanded for choosing to associate with people promoting racist theories. But this isn’t a normal presidency. No one in the DHS is going to criticize Bovino. And, while no one seems all that eager to return Bovino to anti-migration front lines, he’s still going to keep being paid by the US public to cheer on racism from the sidelines, when not traveling overseas to do the same thing from the stage.
Filed Under: bigotry, cbp, gregory bovino, mass deportation, nazis, portugal, trump administration, white nationalism
Companies: resum26
Israel says it has killed new Hamas military leader in Gaza City airstrikes
NASA taps Blue Origin to deliver lunar rovers for Moon Base initiative
2026 NBA Finals schedule, odds: Knicks await Thunder or Spurs after winning East
XRP *JUST* SUCCEEDED!!!! CLARITY ACT EXPOSED!!! (SHE EXPOSED IT)
This is BROKEN! INSANE 5x MONEY CAR WASH WEEK! The NEW GTA Online UPDATE Today! (GTA5 New Update)
Micron Crosses $1 Trillion Market Cap as AI Demand Reshapes Memory Sector
Selena Gomez Reportedly Upset Over Benny Blanco’s Comments on Her ‘Terrible’ Diet
Nikkei 225 Surges Past 65,000 for First Time as Iran Peace Hopes Fuel Record Rally
China assigns ID codes to 28,000+ humanoid robots
Waymo dominates autonomous vehicle registrations as Tesla trails behind
‘Breaking Bad’ Star’s Easy-to-Binge 6-Part Crime Series Spin-Off Is Finally Heading to Free Streaming
SpaceX just won a second Golden Dome contract. This one is $4.16 billion.
The Samsung pay deal is the moment Korean unions changed register
Days of our Lives 2-Week Spoilers May 25-June 5: Gwen Rages, Abe Confesses & 2 Tragic Anniversaries!
SHE IS KILLING XRP!!! WATCH URGENT AND ACT FAST
FIRST NIGHT REVIEW: Take That bring the Circus back to life in spectacular sun-soaked style
Taylor Swift Fans Label Travis Kelce’s Beer-Chugging A ‘Red Flag
Millions of AI agents imperiled by critical vulnerability in open source package
SpaceX’s $2 Trillion IPO: Why Tech Giants Nvidia (NVDA), Apple (AAPL), and Microsoft (MSFT) May Face Pressure
CFTC Has Approved the First Regulated Bitcoin Perpetual Contract in the U.S.
You must be logged in to post a comment Login