Tech News Hub
Meta Launches Llama 4: Smarter AI Models Built for the Next Generation of Tasks
Meta has officially introduced its latest collection of AI models under the Llama 4 banner. The release, which surprisingly happened over the weekend, includes three major models: Llama 4 Scout, Llama 4 Maverick, and the still-in-training Llama 4 Behemoth. Designed to handle complex reasoning across text, images, and videos, these models mark a significant step forward in Meta's open AI strategy.
What’s New in Llama 4
Llama 4 Scout
Scout is built for tasks like document summarization, reasoning over large codebases, and understanding long-form content.
- 109 billion total parameters
- 16 expert models, with 17 billion active parameters
- Massive context window of 10 million tokens
- Efficient enough to run on a single Nvidia H100 GPU
Llama 4 Maverick
Maverick is aimed at general assistant and chatbot use cases, such as creative writing and coding help.
- 400 billion total parameters
- 128 experts, also with 17 billion active parameters
- Requires more computing power, such as an Nvidia H100 DGX system
- Outperforms OpenAI’s GPT-4o and Google’s Gemini 2.0 in certain benchmarks, but not quite at the level of the latest GPT-4.5 or Gemini 2.5 Pro
Llama 4 Behemoth (coming soon)
Behemoth is Meta’s most powerful model yet.
- Nearly 2 trillion total parameters
- 16 experts with 288 billion active parameters
- Outperforms GPT-4.5, Claude 3.7 Sonnet, and Gemini 2.0 Pro in tasks like math and problem-solving
All three models use a Mixture of Experts (MoE) architecture. Instead of having one large model handle everything, MoE allows different parts of the model — called experts — to focus on specific subtasks. This approach improves performance and efficiency, especially when dealing with complex queries.
Availability and Licensing
Scout and Maverick are available now through Llama.com and platforms like Hugging Face. Behemoth is still being trained.
Meta has already integrated Llama 4 into Meta AI, the company’s virtual assistant used across apps like WhatsApp, Messenger, and Instagram in 40 countries. For now, the assistant’s ability to handle image and video input is only available in the United States, and only in English.
However, there are some notable licensing restrictions:
- Developers and companies located in the European Union are not allowed to use or distribute the models, likely due to regional AI and data privacy regulations.
- Companies with more than 700 million monthly active users must request special permission from Meta to use Llama 4.
A More Balanced and Responsive AI
One major update in Llama 4 is how the models respond to politically and socially sensitive questions. Meta says it has tuned the models to:
- Answer a broader range of questions — even controversial ones — more consistently
- Avoid outright refusal unless truly necessary
- Offer factual responses without judgment or favoritism
These changes come in response to criticism that AI models are too selective or biased in their responses. Some political figures have accused AI platforms of promoting certain viewpoints, but Meta’s latest update attempts to make its models more neutral and flexible across different perspectives.
How They Compare
Image Credits:chatgpt
Final Thoughts
Llama 4 is a major step forward for Meta in the competitive AI landscape. With powerful new models, support for multimodal input, and a more open approach to content, Meta is pushing the boundaries of what open-source AI can do.
Despite some limitations in availability and licensing, Llama 4 is likely to attract strong interest from developers, researchers, and companies looking for high-performance AI tools that can be adapted to real-world applications.
Related Stories:
Top Stories:
Anthropic Introduces Web Search for Claude Chatbot with Real-Time Insights
Anthropic has added a web search feature to its AI-powered chatbot, Claude, giving it access to real-time information for improved responses. This long-awaited update aims to enhance Claude’s capabilities and bring it closer to rival chatbots like OpenAI’s ChatGPT and Google’s Gemini.
Web Search Now Available for Paid Users
Web search is currently available in preview for paid Claude users in the U.S., with plans to expand access to free users and other regions soon. Users can enable this feature by navigating to their profile settings in the Claude web app. Once activated, Claude will automatically search the web when needed to deliver more informed responses.
Powered by the Latest Claude 3.7 Sonnet Model
Anthropic’s web search feature is currently supported only by the Claude 3.7 Sonnet model. This ensures the chatbot’s responses integrate the most up-to-date information when required.
Enhanced Responses with Source Citations
When providing information from web searches, Claude includes direct citations to ensure transparency and allow users to verify sources easily. This approach aims to improve credibility and help users distinguish between factual data and potential misinformation.
Potential Limitations
While the web search feature enhances Claude’s knowledge base, early tests reveal occasional inconsistencies. Responses for some current event queries may not always trigger web search, but when they do, Claude effectively cites sources from platforms like NPR, Reuters, and even social media sites such as X.
Competing with Other AI Chatbots
By adding web search, Claude now aligns with competitors like ChatGPT, Gemini, and Mistral’s Le Chat. Anthropic previously resisted adding this feature, believing Claude’s self-contained design was sufficient. However, growing competition likely influenced this change.
Accuracy Concerns
Despite the added functionality, risks remain. Like other chatbots, Claude may still misinterpret or misrepresent information. Studies have shown that popular AI chatbots can provide incorrect answers in over 60% of cases.
Anthropic’s move to introduce web search is a significant step in strengthening Claude’s capabilities, offering users access to broader information with improved accuracy measures.
Related Stories:
Top Stories:
Nvidia Unveils DGX Spark and DGX Station: Bringing AI Supercomputing to Your Desk
Nvidia recently unveiled two mighty AI-benchmarked machines at GTC 2025, which will transport supercomputer-grade performance to your desk. Say hello to the DGX Spark and DGX Station, engineered to perform arduous AI tasks seamlessly.
What Makes These Machines So Unique?
In his keynote address, Nvidia CEO Jensen Huang referred to these new devices as the future computers. "This is what computers ought to look like, and this is what computers will run in the future," he clarified.
These are what distinguish them:
- DGX Spark — With Nvidia's GB10 Grace Blackwell Superchip powering it, this device can handle as much as 1,000 trillion AI operations per second. It's ideal for prototyping, tuning, and executing AI models at the edge.
- DGX Station — Featuring the GB300 Grace Blackwell Ultra Desktop Superchip and a whopping 784GB of memory, this workstation is built to handle heavy-duty AI workloads with serious velocity and efficiency.
When Can You Get One?
- The DGX Spark is ready now for those willing to jump into AI development.
- The DGX Station will ship later this year via premier partners such as Asus, Boxx, Dell, HP, and Lenovo.
Why Does This Matter?
Huang thinks these machines are the key to AI's future. "AI agents will be everywhere. and we need a new line of computers for this new era," he said.
With the DGX Spark and DGX Station, Nvidia is providing developers and businesses with the capabilities they require to advance AI innovation further than ever before.
Related Stories:
Top Stories:
Elon Musk’s xAI Acquires Hotshot to Strengthen AI-Powered Video Generation
Elon Musk’s artificial intelligence company, xAI, has acquired Hotshot, a startup known for developing advanced AI video generation tools. This move signals xAI’s intent to expand its presence in the growing field of generative video technology, potentially competing with industry leaders like OpenAI’s Sora and Google’s Veo 2.
Hotshot’s Journey and Achievements
Based in San Francisco and co-founded by Aakash Sastry and John Mullan, Hotshot began by specializing in AI-driven photo editing before adapting its focus towards text-to-video models. Over the last two years, the firm built three video foundation models:
- Hotshot-XL
- Hotshot Act One
- Hotshot
Hotshot's work by CEO Aakash Sastry has given much insight into the impact AI content generation will have on education, entertainment, communication, and productivity.
Integration with xAI
After the acquisition, Hotshot's technology will further scale with xAI's robust computing power, Colossus, which is one of the world's largest clusters. The deal is in line with Musk's previous indications of xAI's intentions to build video-generation models for its Grok chatbot platform. In a January livestream, Musk said that a "Grok Video" model might be released in a few months.
Impact on Hotshot Users
Hotshot has already begun winding down its services. The company announced that new video creation stopped on March 14, and existing users have until March 30 to download their previously created content.
Uncertain Future for Hotshot’s Team
It remains unclear whether the entire Hotshot team will join xAI. While Sastry declined to comment on this aspect, the acquisition reflects xAI’s aggressive move into the competitive AI video generation space.
With this purchase, xAI is set to take on industry giants in AI-powered content generation, further enhancing its creative prowess in the technology sector.
Related Stories:
Top Stories:
Meta Explores In-House AI Chips to Reduce Reliance on Nvidia
Meta is testing its own custom chip designed to train AI systems, aiming to reduce its dependence on hardware suppliers like Nvidia.
The chip, which was co-developed with Taiwan's TSMC, is being tested on a small scale. If the trials are successful, Meta expects to scale up manufacturing to enable bigger AI training jobs.
This isn't the first time Meta has dabbled in custom chips — the firm initially launched hardware to operate AI models, but not train them. Previous chip initiatives were downsized or canceled after not meeting projections.
At a time when Meta is set to spend as much as $65 billion on infrastructure this year — a large portion of that going towards Nvidia GPUs — succeeding with in-house chips would dramatically reduce expenditures.
By creating its own chips, Meta is moving towards greater control over its AI infrastructure and greater efficiency in handling demanding workloads.
Related Stories:
Top Stories:
Opera Introduces Browser Operator: An AI Agent That Works Natively Within Your Browser
Opera announced Browser Operator, an AI agent that can run tasks natively inside the browser. In contrast to cloud-based AI solutions, this agent operates natively on the device, which makes it faster and more secure.
How Browser Operator Operates
In a demo video, Opera demonstrated Browser Operator doing real-world tasks, such as:
- Online shopping: Searching for a pair of socks on Walmart.
- Booking tickets: Finding football match tickets on an official club website.
- Travel planning: Looking up flights and hotels on Booking.com.
The AI agent can engage with various websites and thus serve as a convenient utility for streamlining everyday online activities. Still, Opera has not publicly asserted that it can process more general queries such as, "Find the cheapest flight from London to New York tomorrow" across multiple sites.
Native AI for Better Security
One of the most attractive aspects of Browser Operator is on-device execution. In contrast to rival AI agents that use cloud-based virtual machines, this ensures more privacy and security. The users will also have complete transparency and control of the agent's behavior, being able to stop it at any moment.
The AI-Powered Browser Race
Opera is getting into a competitive arena where some companies are vying to build AI-powered browser experiences:
- OpenAI launched "Operator," a browser-based ChatGPT Pro AI agent,
- Browser Company developing "Dia" - a novel browser with natively integrated AI-driven automation
- Perplexity AI is building out "Comet," an AI-fueled browser
When Can You Give It a Whirl?
Browser Operator is scheduled to be delivered in the not-too-distant future via the Opera Feature Drop initiative for those willing to experiment early with wthe ider release planned in the works.
As AI is soon changing the way we engage with the web at a lightning-fast pace, Opera's action foreshadows an era where browsers not only present information but also help users get tasks done.
Related Stories:
Top Stories:
OpenAI Plans to Integrate Sora’s Video Generation into ChatGPT
OpenAI is gearing up to integrate its AI video-generation tool, Sora, into ChatGPT, bringing advanced video creation capabilities directly to its popular chatbot. This move was confirmed during an office hours session on Discord, where OpenAI’s product lead for Sora, Rohan Sahai, shared insights into the company’s plans.
Sora’s Expansion Beyond a Standalone Web App
Sora is currently only accessible via a standalone web app that OpenAI introduced in December. It allows users to create up to 20-second cinematic video clips with the help of AI. OpenAI now wants to make Sora more accessible by integrating it into ChatGPT. Sahai did not give a specific timeline, but he suggested that the integration could provide a more streamlined experience than the full-fledged Sora web app, where users can edit and splice videos.
One of the primary reasons OpenAI initially isolated Sora was to preserve ChatGPT's simplicity. But with Sora integrated, ChatGPT could become even more potent, enabling users to produce videos without ever having to leave the chatbot interface.
Potential Benefits of the Integration
By adding Sora to ChatGPT, OpenAI would be able to make more users join and prompt them to upgrade to premium membership levels. The integration can also have varying video generation capacity, with advanced functions available to paid members.
OpenAI has also been improving Sora's web experience, adding means for users to navigate community-created videos. The company is also exploring launching a separate mobile app for Sora and is actively looking for mobile engineers to build it.
Expanding AI Capabilities: Sora-Powered Image Generation
Beyond video, OpenAI is working on expanding Sora’s capabilities to AI-generated images. Sahai confirmed that OpenAI is developing an image generator powered by Sora, which could enhance the realism and quality of generated images. While ChatGPT already supports image generation via OpenAI’s DALL-E 3 model, a Sora-powered image generator may take things a step further by creating highly photorealistic visuals.
What’s Next for Sora?
Aside from the integration of ChatGPT, OpenAI is also working on an enhanced version of Sora Turbo, the model behind the existing Sora web app. The enhancement seeks to enhance video creation capabilities to make Sora even more sophisticated and accessible.
As OpenAI keeps pushing the limits of AI-created content, Sora's integration with ChatGPT might be a giant leap toward simplifying video creation for everyone.
Related Stories:
Top Stories:
