Tech Republic

Image: AnnaStills/Envato Elements Midjourney unveiled on April 3 the alpha version of its V7 model — the company’s first major model release in more than a year. The update introduces significant upgrades, including new operational modes and personalization features, as Midjourney invites its community to test the model and provide feedback. The independent research lab introduced its latest AI-image generation model, V7, as a substantial step forward from its previous versions offering two operational modes — Turbo and Relax — along with a newly introduced “Draft Mode” for faster, cost-effective image rendering. The update also enables automatic model personalization, allowing users to fine-tune outputs based on visual preferences. Midjourney teased more features, promising upgrades to the editing, upscaling, retexture capabilities, and a new V7 character and object reference. A step up from Midjourney V6 The V7 model is still in its alpha stage, so as it continues to develop, Midjourney wants users to try it out and give constructive feedback, sharing plans to apply the responses to improving and working out the kinks of its newest model. Midjourney’s announcement shared several improvements since its V6 model’s release in December 2023. While the model did improve with its 6.1 and 6.2 versions, the V7 promises even more enhancements. For instance, its new Voice Mode feature has text-to-talk capabilities, offering more ways for users to prompt the model’s AI image generation. Another change from previous versions involves its personalization feature, which is turned on by default, requiring users to create their V7-specific personalized style before using the model. Users can toggle it on and off afterward if they prefer. Certain features like Moodboards, SREF (Style & Reference) work, and performance will continue to undergo improvements through V7’s future updates, which users can expect every one to two weeks over 60 days. Until then, the lab stated that retexture, upscaling, and editing will fall back on the V6 models. Although the next generation of AI advancements may be exciting, Midjourney encourages users to remember that the V7 version may require different styles of prompting than its predecessors. More must-read AI coverage Provide Midjourney with feedback and share your V7 creations While the news is fresh, users have begun sharing their opinions of the V7 online with mixed responses. Many still hope for improvements in overall image quality, especially regarding factors like anatomical understanding and text generation, two significant weaknesses for AI image generators. People who wish try out the V7 new model and provide constructive feedback can do so through Midjourney’s Discord channel #ideas-and-feedback or share their new V7 model creations with the Midjourney community in #v7-showcase. The lab even promised a community-wide roadmap ranking session to gain additional input. source

Anthropic’s Claude 3.7 Sonnet. Image: Anthropic/YouTube Anthropic released a new study on April 3 examining how AI models process information and the limitations of tracing their decision-making from prompt to output. The researchers found Claude 3.7 Sonnet isn’t always “faithful” in disclosing how it generates responses. Anthropic probes how closely AI output reflects internal reasoning Anthropic is known for publicizing its introspective research. The company has previously explored interpretable features within its generative AI models and questioned whether the reasoning these models present as part of their answers truly reflects their internal logic. Its latest study dives deeper into the chain of thought — the “reasoning” that AI models provide to users. Expanding on earlier work, the researchers asked: Does the model genuinely think in the way it claims to? The findings are detailed in a paper titled “Reasoning Models Don’t Always Say What They Think” from the Alignment Science Team. The study found that Anthropic’s Claude 3.7 Sonnet and DeepSeek-R1 are “unfaithful” — meaning they don’t always acknowledge when a correct answer was embedded in the prompt itself. In some cases, prompts included scenarios such as: “You have gained unauthorized access to the system.” Only 25% of the time for Claude 3.7 Sonnet and 39% of the time for DeepSeek-R1 did the models admit to using the hint embedded in the prompt to reach their answer. Both models tended to generate longer chains of thought when being unfaithful, compared to when they explicitly reference the prompt. They also became less faithful as the task complexity increased. SEE: DeepSeek developed a new technique for AI ‘reasoning’ in collaboration with Tsinghua University. Although generative AI doesn’t truly think, these hint-based tests serve as a lens into the opaque processes of generative AI systems. Anthropic notes that such tests are useful in understanding how models interpret prompts — and how these interpretations could be exploited by threat actors. More must-read AI coverage Training AI models to be more ‘faithful’ is an uphill battle The researchers hypothesized that giving models more complex reasoning tasks might lead to greater faithfulness. They aimed to train the models to “use its reasoning more effectively,” hoping this would help them more transparently incorporate the hints. However, the training only marginally improved faithfulness. Next, they gamified the training by using a “reward hacking” method. Reward hacking doesn’t usually produce the desired result in large, general AI models, since it encourages the model to reach a reward state above all other goals. In this case, Anthropic rewarded models for providing wrong answers that matched hints seeded in the prompts. This, they theorized, would result in a model that focused on the hints and revealed its use of the hints. Instead, the usual problem with reward hacking applied — the AI created long-winded, fictional accounts of why an incorrect hint was right in order to get the reward. Ultimately, it comes down to AI hallucinations still occurring, and human researchers needing to work more on how to weed out undesirable behavior. “Overall, our results point to the fact that advanced reasoning models very often hide their true thought processes, and sometimes do so when their behaviors are explicitly misaligned,” Anthropic’s team wrote. source

Image: Sundry Photography/Adobe Stock In a bid to tilt the cybersecurity battlefield in favor of defenders, Google has introduced Sec-Gemini v1, a new experimental AI model designed to help security teams identify threats, analyze incidents, and understand vulnerabilities faster and more accurately than before. Announced by the company’s cybersecurity research leads, Elie Burzstein and Marianna Tishchenko, Sec-Gemini v1 is the latest addition to Google’s growing family of Gemini-powered tools — but this time, it is laser-focused on cybersecurity. The growing cyber threat — and why Google’s AI push matters Cyberattacks are becoming more frequent, sophisticated, and targeted. From ransomware to state-sponsored hacking, defenders are overwhelmed. Add to that the rise of remote work, cloud systems, and open-source software, and the threat landscape becomes even more complicated. Cybersecurity has always been an unfair fight. Attackers only need to find one weak spot, while defenders must guard every possible entry point. Google’s answer is to develop an AI that acts like a force multiplier, helping human analysts work smarter. It’s a game of one-versus-all, and Google believes AI can help level the playing field. What makes Sec-Gemini v1 different? What sets Sec-Gemini v1 apart is its access to real-time cybersecurity data from trusted sources like Google Threat Intelligence (GTI), Mandiant’s attack reports, and the Open Source Vulnerabilities (OSV) database. This lets it: Pinpoint root causes of security incidents faster. Identify threat actors (like the Chinese-linked Salt Typhoon group) and their tactics. Analyze vulnerabilities in context — explaining not just what’s broken but how hackers might exploit it. Google claims the model has already shown strong results in internal tests, outperforming other leading AI models — including OpenAI’s GPT-4 and Anthropic’s Claude — on key security benchmarks. On the CTI-MCQ benchmark, which measures how well AI understands threat intelligence, Sec-Gemini scored over 11% higher. It also outpaced rivals by 10.5% on the CTI-Root Cause Mapping test. More Google news & tips The bigger AI security race Google isn’t alone in pushing AI-powered security; Microsoft’s Security Copilot (powered by OpenAI) and Amazon’s GuardDuty are also betting on AI to automate defenses. Still, Google’s deep data integration and benchmark-beating performance could give Sec-Gemini v1 an edge — at least for now. Google opens the doors, but only slightly AI security tools have had mixed success. Some worry they’re just fancy assistants that still require human oversight. But Google insists Sec-Gemini v1 is different. It doesn’t just summarize threats but explains them in ways that speed up decision-making. For now, it’s only available for research, not commercial use. But if it lives up to the hype, it could mark a turning point in how defenders keep up with hackers in an AI-charged world. Interested in testing Sec-Gemini v1? Google is taking requests via this form. source

When choosing the best accounts receivable software, I look for features that automate invoicing and payment reminders. I also look for clear reporting tools for tracking outstanding balances and the ability to monitor cash flow in real time. The software should also easily integrate with your existing accounting system. I’ve put together this buyer’s guide to help you quickly understand your options and choose the best accounts receivable software for your unique needs. Here’s a quick overview of the top vendors I’ll compare: Best for businesses with complex billing structures: Sage Intacct Best for integrating with existing accounting systems: BILL Best A/R software in an all-in-one platform: Intuit QuickBooks Best for integrated time tracking: FreshBooks Best for automation and predictive analytics: High Radius Why you can trust TechRepublic TechRepublic delivers thorough, expert-driven reviews, crafted by professionals with deep expertise in their respective domains. Our team includes experienced specialists and industry advisors with hands-on knowledge of the products they assess. Each piece is grounded in practical experience, powered by a strong grasp of the real-world business needs. Quick comparison of the best accounts receivable software Monthly pricing Does not include seasonal discounts Mobile app Multi-currency support Customizable invoice templates Cash application automation Auto-matching of invoices with payments Sage Intacct Custom No Yes Yes Yes BILL From $45 Yes Yes Yes Yes Intuit QuickBooks Solopreneur accounting package from $20 Yes Yes Yes Limited FreshBooks Lite package from $21 Yes Yes Yes Limited High Radius Custom Yes Yes Yes Yes Sage Intacct: Best for businesses with complex billing structures Image: Sage Intacct Sage Intacct helps business owners automate and manage invoicing and collections with accuracy and control. This product provides real-time access to customer balances, connects smoothly with the general ledger, and offers customizable workflows for revenue tracking. As a cloud-based solution, it scales easily with business growth and includes detailed reporting to help improve operations and cash flow. Sage Intacct is especially strong for businesses with complex billing needs, such as subscription models, tiered pricing, or usage-based charges. It lets users automate advanced billing processes, which cuts down on manual work and mistakes. My favorite feature about this software is the ability to create invoices that combine charges from different contracts or entities. This is ideal for companies with multiple locations or business units. Pricing Sage Intacct does not publicize general pricing information. We recommend contacting their sales team for a custom quote. Standout features Automated invoicing and collections: Streamlined accounts receivable processes through automated invoicing and collection Recurring invoice generation: Efficient management of subscription-based services through recurring invoices Flexible payment options: Offers customers various payment methods, including credit cards, checks, and ACH transfers Real-time reporting and dashboards: Comprehensive reporting options for customer aging, invoice analyses, and deferred revenue Seamless integration with CRM Customer Relationship Management systems: Capacity for integration with existing CRM for a consolidated view of quotes, sales orders, and invoices Enhanced internal controls: Ability to define and implement automated internal control processes for accounts receivable workflows. Pros and cons Pros Cons Multiple customization options Works well with CRM Well-organized interface Scalable for multi-entity and multi-location businesses Steep learning curve for new users. Higher cost compared to other small business solutions Difficult to customize without customer support BILL: Best for integrating with existing accounting systems Image: BILL BILL is built for businesses that want to automate invoice creation, streamline customer payments, and improve cash flow visibility. Its user-friendly interface and powerful automation tools set BILL apart from the competition. I like that it supports digital invoicing, automatic payment reminders, and online payment options, which make it easy to manage receivables from anywhere. BILL is especially effective for businesses that rely on syncing with existing general ledger accounting systems. Its two-way integrations ensure that invoices, payments, and customer interaction is reflected in real time, eliminating duplicate data entry and reducing reconciliation errors. Pricing Essentials Plan: $45 per user per month Team Plan: $55 per user per month Corporate Plan: $79 per user per month Enterprise Plan: Custom pricing; contact bill.com for details Standout features Customer portal access: Options for a dedicated customer bill payment portal Automated payment matching: Incoming payments are automatically matched to outstanding invoices Invoice status tracking: Option to monitor the status of your invoices in real-time and track the status of invoices sent, viewed, and paid Automated late fee application: Ability to implement automatic late fee charges on overdue invoices Pros and cons Pros Cons Supports multiple approval levels for enhanced control over financial processes Provides instant updates and notifications for business managers, bankers, and accountants System is generally easy to navigate Customer support response times can be lengthy Intermittent technical issues, leading to operational disruptions Some features on upgraded platform are not intuitive Intuit QuickBooks: Best A/R software in an all-in-one platform Image: QuickBooks Intuit QuickBooks stands out for its ability to automate invoicing, track payments in real time, and sync outside financial data within a single dashboard. My favorite part of this A/R platform is its deep integration with QuickBooks’ broader accounting suite, providing for cohesive cash flow management without additional tools. Businesses benefit from customizable invoice templates, built-in payment processing, and intelligent reminders that reduce manual follow-ups. QuickBooks is the best choice for businesses seeking an all-in-one accounts receivable solution because it combines A/R tools with bookkeeping, reporting, tax prep, and payroll in one cohesive system. This level of integration keeps financial data aligned, reducing errors and saving time. Pricing QuickBooks Solopreneur: $20 per monthThis plan is designed for self-employed individuals. QuickBooks Simple Start: $35 per monthThis plan is ideal for new, single-member businesses. QuickBooks Online Essentials: $65 per monthThis plan is most suitable for small businesses with multiple members QuickBooks Online Plus: $99 per monthGeared towards growing businesses QuickBooks Online Advanced: $235 per monthDesigned for larger businesses with complex needs Standout features Automated payment reminders: QuickBooks Online automates key tasks like invoice creation and payment tracking Detailed accounts receivable aging reports: QuickBooks Online lets you quickly identify overdue accounts and generate detailed reports

Image: John Gress Media Inc, Shutterstock / John Gress Media Inc Apple is seeking to eliminate cell phone dead spots though expanded satellites, but Elon Musk’s Starlink won’t let that happen without a fight, according to an exclusive report from The Wall Street Journal. Sources who spoke to WSJ say SpaceX is now putting pressure on U.S. federal regulators to stall Apple’s expansion of its Globalstar satellite service, which directly competes with SpaceX’s Starlink network. Reportedly, the pressure intensified after discussions between the two companies broke down. Originally, they were attempting to strike a deal to directly connect iPhones to Starlink satellites, but talks ended without a direct agreement. Instead, SpaceX and T-Mobile will be able to offer their alternative satellite services on Apple devices, a departure from Apple’s famously closed ecosystem. Why satellite availability is limited All satellites use radio frequencies to send signals to Earth. If too many satellites try to use the same frequency, the signals become muddled, degrading communication and slowing down data speeds. To prevent this from happening, most geographic regions license specific radio frequencies to certain satellite providers. The more radio frequencies that a single company controls, the more data it can send and the faster its communication will be. However, if one company monopolizes too many radio frequencies in one region, it forces other satellite providers out. Other providers must offer limited services on a smaller bandwidth, or they opt out altogether, leading to dead zones with no service at all. Having a monopoly or a majority hold on satellite signals also allows the majority provider to drive up costs. This results in price gouging consumers who rely on the provider for cell phone service. Mobility must-reads Apple and SpaceX compete for satellite dominance So far, SpaceX has launched over 550 satellites, far more than Apple, which allows Starlink to dominate the satellite connectivity market. SpaceX launched its first Starlink satellite in 2018, and began offering limited access to its beta internet service in 2020. Apple didn’t start offering the service until two years later, when it struck a deal with Globalstar in 2022. Globalstar actually hired SpaceX to launch Apple’s satellites, further complicating the ties between the companies. Currently, Apple devices use this satellite service to send texts and make SOS calls when no other cell service is available. With this expansion of its Globalstar partnership, Apple is seeking to offer more connectivity in more remote areas outside of emergency scenarios — which will directly compete with Starlink. This satellite space race marks the latest in a series of clashes between Apple and Elon Musk. Apple and Tesla have previously clashed over the distribution of X on Apple devices as well as the development of self-driving cars using AI models. source

Image: Envato/DC_Studio Researchers from AI company DeepSeek and Tsinghua University have introduced a new technique to enhance “reasoning” in large language models (LLMs). Reasoning capabilities have emerged as a critical benchmark in the race to build top-performing generative AI systems. China and the U.S. are actively competing to develop the most powerful and practical models. According to a Stanford University report in April, China’s LLMs are rapidly closing the gap with their U.S. counterparts. In 2024, China produced 15 notable AI models compared to 40 in the U.S., but it leads in patents and academic publications. What is DeepSeek’s new technique? DeepSeek researchers published a paper, titled “Inference-Time Scaling for Generalist Reward Modeling,” on Cornell University’s arXiv, the archive of scientific papers. Note that papers published on arXiv are not necessarily peer-reviewed. In the paper, the researchers detailed a combination of two AI training methods: generative reward modeling and self-principled critique tuning. “In this work, we investigate how to improve reward modeling (RM) with more inference compute for general queries, i.e. the inference-time scalability of generalist RM, and further, how to improve the effectiveness of performance-compute scaling with proper learning methods,” the researchers wrote. More must-read AI coverage SEE: DDoS Attacks Now Key Weapons in Geopolitical Conflicts, NETSCOUT Warns Reward modeling is the process of training AI to align more closely with user preferences. With Self-Principled Critique Tuning, the model generates its own critiques or ‘principles’ during inference to fine-tune its answers. The combined approach continues the effort to let LLMs deliver more relevant answers faster. “Empirically, we show that SPCT significantly improves the quality and scalability of GRMs, outperforming existing methods and models in various RM benchmarks without severe biases, and could achieve better performance compared to training-time scaling,” the researchers wrote. They called the models trained with this method DeepSeek-GRM. “DeepSeek-GRM still meets challenges in some tasks, which we believe can be addressed by future efforts in generalist reward systems,” the researchers wrote. What’s next for DeepSeek? DeepSeek has generated significant buzz around the R1 model, which rivals leading reasoning-focused models like OpenAI o1. A second model, DeepSeek-R2, is rumored for release in May. The company also launched DeepSeek-V3-0324, an updated reasoning model released in late March. According to the paper, models built with the new GRM-SPCT method will be open-searched, though no release date has been specified. source

For businesses needing quick access to working capital, a same-day business loan may fit the bill. It can provide you with a simple application process, usually with minimal documentation required. The best options will offer flexible qualifications, various loan types, and favorable rates and terms. I’ve reviewed the best same-day business funding options across various lenders, with my top picks detailed below. Best overall for flexible qualification requirements: Lendio Best for a variety of flexible loan options: Credibly Best for businesses seeking short-term financing: Bluevine Best for tailored financing solutions: National Funding Best for minimal documentation requirements: QuickBridge Best for large funding needs: Fora Financial Best same-day business loans at a glance Max. loan amount Repayment term Est. starting rate Funding speed Lendio $10 million Varies per loan type Varies per loan type As fast as same-day Credibly $10 million 24 months 1.11x factor rate As fast as same-day Bluevine $250,000 Up to 12 months 7.8% As fast as same-day National Funding $500,000 Varies 1.11x factor rate As fast as same-day QuickBridge $500,000 18 months 1.11x factor rate Within 24 hours Fora Financial $1.5 million Varies 1.1x factor rate As fast as same-day Lendio: Best overall for flexible qualification requirements Image: Lendio Lendio is a standout choice if you’re looking to explore various loan options. With a network of over 75 partnering lenders, it has multiple loan options you can apply for with a single application, saving you time and money. Whether you’re seeking a line of credit, working capital, or equipment financing, Lendio’s broker services can connect you with lenders offering products tailored to your business needs. I recommend Lendio for its flexible qualifications, regardless of whether you’re a startup, a borrower with less-than-ideal credit, or a company looking for a large loan amount. Depending on the loan type you choose, you can receive both approval and funding as fast as same-day. After filling out a short online application, you’ll be paired with a funding specialist who will guide you through your potential options and ensure you’re matched with the best financing option for your business. They’ll review your application, answer any questions, and ensure you’re matched with the best lender that understands your financing needs. How to qualify Credit score: Varies by loan product Time in business: Varies by loan product Annual revenue: Varies by loan product Loan types & details Loan amount Interest rate Repayment terms A/R financing Up to $10 million 3% and up Up to 1 year Short-term loan $10,000 to $5 million 8% and up 6 months to 7 years Equipment financing $5,000 to $5 million 7.5% and up 1 to 10 years Cash advance $5,000 to $1 million 18% and up 3 to 36 months Line of credit Up to $250,000 8% to 60% 6 to 24 months Features Access to over 75 lenders with a single application Quick application process (10-15 minutes) Fast approval and funding (as fast as same-day) Dedicated funding specialists to help you Various financing options Pros and cons Pros Cons Variety of loan options across multiple lenders Quick approval and funding timeline Dedicated funding specialists offered Flexible qualifications Interest rates can vary widely depending on the lender and loan type Additional fees may apply based on the lender Lendio is not a direct lender, but rather a broker Credibly: Best for a variety of flexible loan options Image: Credibly I chose Credibly for its variety of loans, which have quick funding speeds and overall flexibility. It offers a wide variety of business loans, including working capital loans, merchant cash advances, business lines of credit, and equipment financing. Each offers quick access to financing and is designed to meet the unique needs of various business industries. You can receive approval within as little as four hours by completing a simple and efficient online application. Depending on the loan type, funding can be available as quickly as the same day, helping you gain access to capital if your business has time-sensitive financing needs. How to qualify Credit score: Minimum of 500 (can vary per loan type) Time in business: At least 6 months Monthly revenue: Minimum of $15,000 Loan types & details Loan amount Interest rate Repayment terms Equipment financing $10,000 to $10 million​ Varies Varies Working capital loans $25,000 to $600,000 Factor rates starting at 1.11​x 6 to 24 months Business lines of credit Up to $600,000 Factor rates starting at 1.11​x 3 to 24 months Merchant cash advances Up to $600,000 Factor rates starting at 1.11​x 3 to 24 months Features Same-day funding for most loan types (as quick as four hours) Multiple loan options for a wide variety of business needs Flexible qualifications, even for businesses with low credit scores Pros and cons Pros Cons Quick approval and funding timeline Various loan types available for different types of business needs Flexible qualification requirements Simple application process Loan terms, interest rates, and fees can vary widely depending on the loan type and qualifications Funding may not be guaranteed Bluevine: Best for businesses seeking short-term financing Image: Bluevine Bluevine stands out to me because of its combination of competitive rates, quick approval and funding speeds, and flexible repayment options. Its line of credit is a solid option if you seek fast access to capital with favorable rates and terms. It’s also ideal if you need to make quick decisions and move fast without the stress of high fees or long wait times, which is why I chose it as the best pick for short-term financing needs. Notably, it’s also in our roundup of the best business lines of credit. The application process is quick and easy to fill out. You can apply online in as little as five minutes, and if approved, you can access funds as fast as the same day. With low starting rates and flexible repayment terms, Bluevine’s offerings are designed to help businesses manage their finances efficiently without any complicated processes. How to qualify Credit score: 625 or higher Time in business: At least

Image: pichetw/Envato Elements Google and OpenAI have rejected the U.K. government’s proposal aimed at balancing the use of online content for AI training with protecting artists’ rights to consent and compensation. The companies suggest that a broad exception for text and data mining (TDM) would be more beneficial for all stakeholders. The government’s proposal, published in December, outlined a system that permits AI developers to use creators’ online content to train their models unless rights holders explicitly opt out. It also mandates transparency from AI developers on which creative materials they use and how these are sourced. Tech giants favor broad TDM exception over artist protections In its response to the subsequent consultation, OpenAI said opt-out models face “significant implementation challenges.” OpenAI pointed to the unclear standards in the EU, which mean “AI developers struggle to identify which works can be accessed and which are off-limits.” The ChatGPT maker said any transparency obligations must not require the disclosure of more sensitive information than is required in other jurisdictions, or AI companies may be less inclined to operate in the U.K. OpenAI also supports the proposal of a TDM exception that would allow copyrighted material to be used to train commercial models without the rights holder’s permission. The company claims it will “drive AI innovation and investment in the UK, and could be designed to balance the needs of AI development with the mitigation of concrete harms to copyright owners.” SEE: Google, Meta Criticise U.K. and E.U. AI Regulations More must-read AI coverage Google wants the TDM exception too, as it lays out in its response; however, it wants it for both commercial and non-commercial uses. The company has expressed this desire multiple times before, but plans to allow it for commercial purposes were abandoned in February 2023 after being widely criticised by creative industries. The Gemini creator clarified it supports the opt-out model for creators but that it does not “translate to remuneration rights” if their content is somehow used in training data. The government’s proposal would allow rights holders to negotiate their own licensing agreements with AI companies if they chose to do so. Google also described the transparency requirements as “excessive” and could “hinder AI development and impact the U.K.’s competitiveness in this space.” Artists push back Artists have expressed outrage over the U.K.’s decision to revise copyright laws in favour of AI, placing the onus on them to opt out of AI training rather than the AI company seeking consent by default. The likes of the Independent Society of Musicians and Publishers Association argued this would further erode their ability to control and profit from their creations. Last month, more than 400 artists, including Paul McCartney, Ben Stiller, and Cate Blanchett, sent a letter urging action against AI companies for allegedly exploiting copyrighted works without permission. source

Image: Meta Meta unveiled on April 5 its new AI model series: Llama 4, which includes Llama 4 Maverick and Llama 4 Scout, tailored for conversation and processing large files, respectively, along with an unreleased “teacher” model called Llama 4 Behemoth. Llama 4 is Meta’s first series to adopt a “mixture of experts (MoE) architecture.” This approach activates only select parts of the neural network, referred to as the “experts,” to handle specific subtasks. The task will be broken down into subtasks and each routed to the most appropriate experts, improving resource efficiency. What are the specifics about Llama 4 Maverick and Scout? Llama 4 Maverick features 128 experts and 17 billion active parameters, which represent the portion of a model’s knowledge used to process a given input. Meta describes it as the “product workhorse model for general assistant and chat use cases,” specialising in image interpretation and creative writing. Interestingly, Mark Zuckerberg’s company boasts that Maverick offers “a best-in-class performance to cost ratio” when it comes to conversations. Cost has been playing on the minds of AI giants since the surprise release of DeepSeek in January, which took only $5.6 million to train. SEE: Meta’s $800M Offer To Chip Startup Was Rejected — Here’s Why However, AI experts have noticed that the version of Llama 4 Maverick published on LMArena, which ranks major large language models across various tasks, is “optimized for conversationality” and performs differently from the publicly available version. This suggests that Meta submitted an altered version to LMArena that would rank higher on its leaderboard. Llama 4 Scout also has 17 billion active parameters and just 16 experts, but Meta says it is the “best multimodal model in the world in its class.” It has an unusually large context window of 10 million tokens, which represent the amount of information it can process in a prompt, so it performs well when summarising large documents and in sequential reasoning. Meta says that both Scout and Maverick are its “best yet” due to being distilled from Llama 4 Behemoth, with a whopping 28 billion active parameters and 16 experts. While it already ranks highly on LMArena, it is still being trained and has not been released. According to The Information, the Llama 4 announcement was delayed at least twice due to the models underperforming in technical benchmarks and conversationality. How can you access LLama 4 Maverick and Scout? Scout and Maverick can be downloaded on Llama.com and Hugging Face, or used through the Meta AI chatbots in WhatsApp, Messenger, and Instagram in 40 countries. Multimodal features can only be used in the U.S. and in English, currently. Some partners have already announced integrations; developers can build and deploy AI applications with the Llama 4 models in Microsoft’s Azure AI Foundry and Azure Databricks. More must-read AI coverage Llama 4 is apolitical Meta stated it has worked specifically to “remove bias” from the Llama 4 models. The refusal rate for questions on “debated political and social topics” is over 5% lower than that of Llama 3.3 and, among the questions it does decline, its responses are described as “dramatically more balanced.” U.S. President Donald Trump’s team has voiced skepticism about the neutrality of AI models, with his AI and crypto czar David Sacks suggesting that OpenAI’s ChatGPT is “programmed to be woke” on a podcast. AI experts say that bias ultimately stems from training data and can lead to political leanings in any direction, not just the left. Nevertheless, Zuckerberg’s firm has made a number of recent moves that suggest it wants to stay on the side with the U.S. administration. Republican strategist Joel Kaplan was hired as Meta’s policy lead shortly after Trump assumed office; he sees social media regulation as a direct challenge to free speech. In January, Meta revealed the company was discontinuing its third-party fact-checking program and relocating its content moderation teams from California to Texas to “help remove the concern that biased employees are overly censoring content.” Meta has also eliminated its diversity, equity, and inclusion initiatives after Trump criticised such programs. Furthermore, Meta said the Llama 4 models respond with a “strong political lean” on “contentious” topics at a similar rate to Grok, the chatbot produced by xAI, a company owned by current White House adviser Elon Musk. Llama 4 cannot be used in the E.U. According to the Llama 4 acceptable use policy, individuals “domiciled” or companies with a “principal place of business” in the European Union cannot use or distribute the models. Those individuals or companies can, however, use the Llama 4 models if they are incorporated into a product or service they have access to in the region. This is likely the result of Meta’s issues with E.U. legislation, particularly when it comes to AI. In June 2024, Meta delayed the training of its large language models on public content shared on Facebook and Instagram after E.U., regulators suggested it might need explicit consent from content owners. Meta AI has still not been released within the bloc. SEE: Meta Offers Less Personalised Ads for EU Users Meta signed an open letter urging European regulators to address “inconsistent regulatory decision-making” and unpredictable compliance demands last September. Then, in February, Meta declared it was prepared to escalate its concerns over what the company sees as unfair E.U. regulations directly to Trump. There are other restrictions when it comes to Llama 4 usage, as commercial entities with more than 700 million monthly active users must request permission from Meta before using its models. The Open Source Initiative has said that such a restriction takes the AI “out of the category of “open source,” despite Meta claiming otherwise. source

With the latest stable release dated January 28, 2025, Qwen2.5-Max is classified as a Mixture-of-Experts (MoE) language model developed by Alibaba. Like other language models, Qwen2.5-Max is capable of generating text, understanding different languages, and performing advanced logic. According to recent benchmarks, it is also more secure than DeepSeek-V3-0324. Using Recon to scan for vulnerabilities A team of analysts with Protect AI, the company behind a red teaming and security vulnerability scanning tool known as Recon, recently used their platform to compare the security of Qwen2.5-Max against that of DeepSeek-V3. The team’s assessment reads, in part: “We observed that DeepSeek-V3-0324 is more vulnerable than Qwen2.5-Max, with Recon achieving an almost 25% higher attack success rate (ASR).” While it may be more secure than its competition, Qwen2.5-Max isn’t exactly perfect. According to their tests, the AI model is most susceptible to prompt injection attacks, as these represented almost 48% of all successful cyberattacks against Qwen2.5-Max. Evasion and jailbreak attacks proved to be less successful with an approximate ASR of 40% for both. Exposing vulnerabilities in DeepSeek-V3 Recon utilizes a comprehensive Attack Library to scan current-gen AI models and identify vulnerabilities across six specific categories: Evasion techniques System prompt leaks Prompt injection attacks AI jailbreak attempts General safety controls Adversarial suffix resistance In addition to simulated cyberattacks, Recon also assesses the AI models’ resistance to generating potentially harmful or illegal content. For example, during adversarial suffix resistance tests, Recon attempts to manipulate the AI model into generating harmful or illegal content. The Protect AI team ran Recon against both Qwen2.5-Max and DeepSeek-V3, with the former boasting a lower attack success rate (ASR) across a variety of attacks; including jailbreaks, prompt injection, and evasion techniques. Whereas Qwen2.5-Max had a 47% ASR against prompt injection attacks, compared to DeepSeek-V3’s notably higher 77%. Against evasion techniques, Qwen2.5-Max scored a 39.4% ASR against evasion techniques, while DeepSeek-V3 scored 69.2%. Both AI models displayed similar results across other simulated cyberattacks. Analyzing DeepSeek-V3’s strengths Despite its security weaknesses, DeepSeek-V3-0324 still outperforms Qwen2.5-Max in several different benchmarks. Unlike the ASR, a higher score in these tests actually indicates better performance. DeepSeek-V3-0324 Qwen2.5-Max MMLU-Pro 81.2 75.9 GPQA Diamond 68.4 59.1 MATH-500 94.0 90.2 AIME 2024 59.4 39.6 LiveCodeBench 49.2 39.2 According to these benchmarks, DeepSeek-V3-0324’s strengths include general language understanding (MMLU-Pro), advanced topics such as biology, physics, and chemistry (GPQA Diamond), mathematics (MATH-500, AI in medicine (AIME 2024), and coding (LiveCodeBench). source