Digital Phablet

Tag: open-source

  • Meta’s New Open Source AI Models Take On GPT, Gemini, Claude

    Meta’s New Open Source AI Models Take On GPT, Gemini, Claude

    Meta has unveiled its newest version of the open-source AI family, known as Llama 4, amidst growing competition in the generative AI sector.

    This latest lineup comprises four models, specifically Llama 4 Scout, Llama 4 Maverick, and Llama 4 Behemoth. According to information shared on Meta’s AI website, these models were trained using extensive datasets of unlabeled text, images, and videos, showcasing their diverse multimodal capabilities.

    As of this past Saturday, the Llama 4 Scout and Llama 4 Maverick models are accessible to users across various Meta platforms, including WhatsApp, Messenger, and Instagram Direct, as well as on Meta’s dedicated AI site, Llama.com. Developers can also find these AI models available in open-source repositories like Hugging Face. The Llama 4 Behemoth model, however, remains in training and has not yet been released. Meta has indicated that the Behemoth model is expected to surpass its counterparts and serve a pivotal role in guiding the other models within the Llama 4 series.

    While internally testing the Llama 4 models, Meta conducted comparisons against competitor AI technologies to assess their capabilities and ideal applications. The company highlighted that Llama 4 Maverick excels in creative writing, outperforming models such as OpenAI’s GPT-4o and Google’s Gemini 2.0 in areas like coding, reasoning, multilingual comprehension, long-context processing, and image generation. However, Maverick faced challenges matching the performance of newer models like Gemini 2.5 Pro, GPT-4.5, and Anthropic Claude 3.7 Sonnet.

    Despite Meta’s assertions that the Behemoth model will outstrip most models—including Gemini 2.5 Pro—the company is still facing challenges in minimizing the hardware costs associated with training its most formidable model.

    TechCrunch has observed that the Chinese AI firm DeepSeek has gained significant traction with its competitively priced models, prompting Meta to closely examine how the rival company managed to develop impactful models like R1 and V3 at lower operational costs than previous iterations of Llama.

    Notably, the Llama 4 Scout model is capable of operating on a single Nvidia H100 GPU, while the Llama 4 Maverick model requires a Nvidia H100 DGX graphics system to function.

    Meta plans to host its inaugural LlamaCon AI conference on April 29. The company also intends to launch a standalone Meta AI chatbot within the second quarter of the year.

    In a parallel development, OpenAI has adjusted its GPT-5 model timeline, with CEO Sam Altman announcing on social media that users should anticipate new reasoning models (o3 and o4-mini) in the upcoming weeks as alternatives to GPT-5. Altman confirmed that GPT-5 will be released in the upcoming months, allowing OpenAI more time to refine the model.

  • GIMP Returns After Seven Years As An Open Source Image Editor

    GIMP Returns After Seven Years As An Open Source Image Editor

    After seven years of dedicated development, the GIMP team proudly announces the release of GIMP 3.0, introducing significant enhancements to this beloved open-source image editing software.

    A key feature of this release is the shift to the GTK3 graphical user interface library, which replaces the now-obsolete GTK2. This transition is expected to improve overall performance and offers users a more modern and responsive interface. Additionally, GIMP 3.0 now incorporates non-destructive editing for many popular filters, enabling users to see adjustments in real-time on their canvas.

    In response to user feedback, GIMP 3.0 includes the ability to select multiple layers at once. This feature simplifies complex editing tasks, allowing users to move, transform, or apply effects to several layers simultaneously.

    The update brings important advancements in text management as well. Users can now add outlines, shadows, bevels, and various styles to text while still maintaining the ability to modify the content, change fonts, and resize. This flexibility is especially useful for designers looking to produce professional-grade text components in their projects.

    The introduction of Wayland support on Linux systems enhances compatibility and performance, offering better integration with modern display servers for a smoother user experience. Furthermore, GIMP 3.0 improves HiDPI support for crisper visuals on high-resolution screens and enhances compatibility with Wacom tablets, which is beneficial for digital artists who depend on precise input methods.

    Curves Non-destructive filter being applied to a portrait
    GIMP

    The plugin ecosystem has greatly expanded as well. GIMP 3.0 adds support for extensions written in Python 3, JavaScript, Lua, and Vala, granting developers more freedom to enhance the software’s capabilities. File format compatibility has also been widened, facilitating easier file exchanges with a broader array of applications. Notably, there is new support for BC7 DDS files and enhanced PSD export capabilities, making cross-platform collaboration simpler. The GIMP development team has also prioritized advancements in color management.

    GIMP 3.0 is now available for GNU/Linux, macOS, and Windows and can be downloaded directly from the official GIMP website. The team aims to accelerate the release schedule for the 3.X series, with GIMP 3.2 anticipated within the next year. Their plan focuses on delivering updates more frequently, even if each version has fewer new features, thus ensuring users can access improvements and refinements more quickly.

  • Meta’s ‘Llamacon’ Event Focuses on Open-Source AI

    Meta’s ‘Llamacon’ Event Focuses on Open-Source AI

    A silhouetted individual holding a smartphone featuring the Facebook logo in front of a sign showcasing the Meta logo.
    SOPA Images / Getty Images

    Meta announced on Tuesday that it will host a new conference for developers called “Llamacon” in April, focusing on advancements in “open source AI technology.”

    Set for April 29, 2025, this event follows the remarkable progress and interest surrounding the company’s open-source Llama models and tools, as stated in Meta’s announcement. Further details, including the event’s location and ticket pricing, have not yet been disclosed, but Meta assures that more information will be revealed soon.

    In addition, Meta announced the return of its Meta Connect event, aimed at professionals in virtual and mixed reality, which is scheduled for September 17-18, 2025. This year’s conference will unveil the latest updates for Meta Horizon and promises to provide insights into the future of technology.

    Like other American tech companies, Meta plans to make significant investments in AI in 2025. At the end of January, Meta’s CEO, Mark Zuckerberg announced that the organization would allocate between $60 billion and $65 billion towards AI infrastructure, which includes a new data center expected to consume one gigawatt of energy—equivalent to the output of two nuclear power stations.

    Zuckerberg remarked, “This will be a pivotal year for AI. I predict that in 2025, Meta AI will serve over one billion users, Llama 4 will be the cutting-edge model, and we will develop an AI engineer capable of contributing significantly to our research and development efforts.”







  • Is Meta’s AI Strategy Working? 6M Downloads vs. 15.5M

    Is Meta’s AI Strategy Working? 6M Downloads vs. 15.5M

    The Landscape of Open Source AI: A Look at Downloads and Popularity

    Open-source AI has taken the tech world by storm, primarily through the release of various large language models (LLMs). With giants like OpenAI and Meta at the forefront, the comparative download statistics reveal insights into market preferences and the impact these models have had on the AI community.

    The Reign of OpenAI’s GPT-2

    OpenAI’s GPT-2 model, released in 2019, has proven to be a game changer for text generation. With a staggering 15.5 million downloads in just a single month, GPT-2 remains the most downloaded model on the HuggingFace repository. This immense popularity can be attributed to its effective training methodology, which focused on an interesting dataset derived from “all the web pages from outbound links on Reddit that received at least 3 karma.” This strategic selection excluded more conventional sources like Wikipedia, making the training data unique among its peers.

    The Essence of GPT-2’s Training Data

    The training data for GPT-2 is particularly intriguing due to its reliance on community-curated Reddit content, showcasing a model that adapts to natural language as communicated in everyday discussions. This could explain the model’s ability to generate human-like text that resonates with users across various domains.

    Meta’s AI Contributions: A Diverse Portfolio

    While OpenAI’s models are leading in popularity, Meta (previously Facebook) also plays a significant role in the open-source AI movement. Their offerings reflect a mix of older and newer model iterations.

    Notable Models from Meta

    1. OPT-125M
      Released in the summer of 2022, this model has achieved 6 million downloads in the last month alone. However, it is still credited under the Facebook label on HuggingFace, indicating the slow transition of branding as Meta repositions its identity within the tech sphere.

    2. Llama 3.1
      This model, although older, continues to be a strong contender with 5.8 million downloads. The Llama series represents Meta’s investment in AI research, showcasing their commitment to open-source methodologies.

    3. Llama 3.3
      The latest iteration, Llama 3.3, has seen 597,000 downloads this past month. While this figure may appear lower compared to its predecessors, it is important to consider the time frame since its release and the competitive landscape of LLMs available to users.

    Other Noteworthy Entrants in the Market

    The domain of open-source AI is not solely dominated by OpenAI and Meta. Several other models have gained remarkable traction:

    • MistralAI’s Nemo Instruct Model

      • Achieved 1.5 million downloads, showcasing an emerging competitor in the LLM arena.
    • Apple’s OpenELM 1.1B Instruct Model
      • Garnered 1.4 million downloads, indicating Apple’s growing interest in the open-source AI sector.

    The Functionality and Challenges of Open Source Models

    A defining feature of open-source models is the ability for anyone to download, adapt, and modify them based on specific licenses. This factor democratizes AI, allowing for exploration and innovation beyond the corporate confines typical for proprietary models like ChatGPT.

    Transparency and Data Usage Challenges

    Despite their accessibility, even established models like Llama exhibit challenges concerning the transparency of their training data. Users often find it difficult to ascertain the datasets utilized, which can complicate effective application and necessitates additional software and knowledge for practical use.

    Balancing the Open-Source Ecosystem

    As the landscape evolves with new models and competitors entering the fray, open source AI continues to enchant developers, researchers, and businesses alike by providing a platform for experimentation and innovation. The contrasting structures of open-source models against proprietary applications like ChatGPT will influence the future of AI development and deployment.

  • This Open-Source ChatGPT Alternative Just Got Real

    This Open-Source ChatGPT Alternative Just Got Real

    French AI startup Mistral made an exciting announcement on Monday, unveiling several new features for its free generative AI assistant, known as le Chat. This update is designed to enhance the chatbot’s capabilities, bringing it in line with leading models from companies like OpenAI and Anthropic.

    One of the key improvements is le Chat’s new ability to search the web and provide citations, similar to the functions offered by platforms like Perplexity and SearchGPT. Additionally, the chatbot has introduced a Canvas feature that allows users to edit and modify content and code, reminiscent of Claude’s Artifacts. Le Chat can now also create images, thanks to its partnership with Black Forest Labs’ Flux Pro, which is the same technology that enables Grok-2’s image generation.

    Furthermore, le Chat has gained the capability to analyze and summarize large PDF documents, including graphs and equations. Mistral is also launching AI “agents” that can streamline repetitive workflows, which employees within a company can share, marking a significant step in automation.

    Many of these advancements are powered by Mistral’s latest model, the Pixtral Large, which boasts 128 billion parameters. Built on the foundation of Mistral Large 2, this frontier model supports a remarkable 128k prompt window and demonstrates outstanding performance across various industry benchmarks, such as MathVista, DocVQA, and VQAv2.

    In its announcement, Mistral highlighted that “Particularly, Pixtral Large is able to understand documents, charts, and natural images.” Additionally, the company is rolling out an updated version of its flagship Mistral Large model, known as Version 24.11, which reportedly shows enhanced performance for tasks necessitating extensive context, such as document analysis.

    Mistral emphasizes its unique approach to artificial intelligence, stating, “We’re not chasing artificial general intelligence at all costs; our mission is to instead place frontier AI in your hands, so you get to decide what to do with advanced AI capabilities.” This strategy has allowed Mistral to manage its resources effectively while consistently providing cutting-edge capabilities at accessible price points. Consequently, all newly introduced features are currently available for free to all users of le Chat.

    Founded in April 2023 by ex-employees from Meta Platforms and Google DeepMind, Mistral is currently valued at $2 billion and offers nearly a dozen different AI models for both commercial and research use. Among these models, Mistral 7B, Mixtral 8x7B, and Mixtral 8x22B are open-source and accessible to the public via Hugging Face, while Mistral’s Small, Medium, and Large models remain closed-source and require the Mistral API for access. Users can access both Large 24.11 and Pixtral Large either through a commercial use license or a more restrictive research license.