Rubbletag

Alibaba's Qwen team has introduced Qwen-Music, an AI music generation model designed to create complete songs with vocals and reinterpret existing music in different styles. The model supports two main tasks: Text to Music Generation and Cover Song Generation. It can create new songs from text descriptions, lyrics and musical attributes, while also reworking songs with different styles and vocal characteristics. Qwen-Music combines three core components: Qwen-Music-Tokenizer, Qwen-Music-LLM and Qwen-Music-Render. Together, these systems process musical information, plan song structure and melodies, and generate high-fidelity stereo audio. A key feature is Melody-CoT, a melody-based chain-of-thought mechanism that allows the model to plan melodies before generating a complete song. According to the technical report, this approach improves creativity, musicality and structural coherence. The model was trained on more than five million hours of multilingual music data covering hundreds of languages, giving it broad exposure to different musical and vocal styles. Qwen-Music is also designed for cover song generation, where it can preserve the melody of reference music while changing style and vocal characteristics. According to its researchers, Qwen-Music achieved leading results across several musicality and audio-quality measurements, with professional evaluators also preferring its outputs over leading proprietary systems in their evaluations. The development reflects Qwen's continued expansion beyond text-based artificial intelligence into specialised generative models capable of working with music, vocals and complex audio creation.

Mistral AI has introduced Robostral Navigate, its first AI model built specifically for embodied navigation, marking the company's move into physical AI and robotic systems. The 8-billion-parameter model is designed to help robots navigate complex environments using only a single RGB camera, without depending on more elaborate multi-sensor systems. Robostral Navigate can process visual information from its surroundings and determine how a robot should move through an environment. Mistral says the model achieved a 76.6% success rate on unseen R2R-CE navigation benchmarks, outperforming several multi-sensor approaches. The model was developed entirely in-house using simulated training data and token-efficient techniques. According to Mistral, it can generalise across different robot types and adapt to real-world obstacles that were not present during training. Robostral Navigate is aimed at embodied AI applications where artificial intelligence must interact with and move through physical environments rather than operate only through text or digital interfaces. The launch reflects Mistral AI's growing interest in physical AI, robotics and industrial applications, extending the company's work beyond language and document intelligence into systems capable of navigating the real world.

Meta's Llama 4 represents a major step in the development of open-weight artificial intelligence, introducing natively multimodal models designed to work across different forms of information. The Llama 4 family includes Scout and Maverick, Meta's first open-weight, natively multimodal models and the first Llama models built using a mixture-of-experts architecture. Llama 4 Scout is designed with an exceptionally large context window of up to 10 million tokens. This allows the model to work with very large amounts of information and makes it suitable for tasks involving extensive documents, large codebases and complex datasets. Llama 4 Maverick focuses on strong multimodal and multilingual capabilities, supporting a wider range of AI applications that require understanding and working across different types of content. The mixture-of-experts architecture allows only selected parts of the model to become active for a particular task. This approach is designed to use computing resources more efficiently while maintaining strong model capabilities. Meta has positioned Llama as an open-weight AI family that gives researchers, developers and businesses greater flexibility to build, adapt and deploy AI systems for their own applications. With Llama 4, Meta continues to advance its open AI strategy by combining multimodal intelligence, large-context capabilities and a more efficient model architecture.

Google is expanding Gemini in Chrome to users in the United Kingdom, bringing more of its AI-powered browser capabilities to a wider audience. Gemini in Chrome is designed to understand the context of the pages and tabs a user is viewing, allowing people to ask questions and receive assistance based on information already open in their browser. The AI can work across multiple tabs, helping users compare information, understand web content and complete research without repeatedly copying and pasting details between different pages. Google is also bringing more agentic capabilities into Chrome. These features are designed to help users carry out multi-step tasks while keeping the user in control of important actions. The expansion reflects Google's broader effort to integrate Gemini directly into everyday browsing, moving AI assistance closer to the websites, information and workflows people already use. By placing Gemini inside Chrome, Google is continuing to develop the browser as a more intelligent workspace where AI can understand context and assist users during everyday online activities.

xAI has expanded Grok's voice capabilities with the introduction of 21 new flagship voices, significantly increasing the range of voice options available across its AI ecosystem. The new voices are multilingual and designed to support more natural, expressive and engaging voice interactions. With this expansion, users and developers have a wider selection of voices for different conversational experiences and applications. The voices are available through xAI's Realtime Voice Agent API, Text to Speech API and Voice Agent Builder, giving developers multiple ways to integrate Grok's voice technology into their products and AI experiences. Alongside the new additions, xAI has also improved its original five flagship voices, focusing on greater naturalness and a more refined conversational experience. The expansion reflects xAI's continued development of voice-based AI interactions, giving users and developers greater flexibility and more choice when building or using conversational AI experiences.

Anthropic has introduced "Reflect", a new Claude feature designed to help users better understand and review how they interact with AI. "Reflect" provides users with insights into their Claude usage patterns, allowing them to look back at how they have been using the AI and consider whether their habits and workflows are helping them stay focused and productive. The feature can provide a monthly recap and includes focus-related controls such as break reminders and quiet hours. These tools are intended to give users greater awareness and control over their AI usage. Rather than focusing only on what Claude can produce, "Reflect" brings attention to the user's own interaction with AI. It encourages people to examine their usage patterns and make adjustments where needed. With "Reflect", Anthropic is adding a more user-focused layer to the Claude experience, combining AI assistance with tools designed to support healthier and more deliberate usage habits.

OpenAI has introduced GPT-5.6, its latest family of frontier AI models, designed to deliver greater intelligence from every token, stronger performance for the cost, and more computing capability for demanding work. The GPT-5.6 family includes Sol, Terra and Luna, giving users and developers different levels of capability, speed and efficiency depending on the complexity of their tasks. GPT-5.6 Sol is OpenAI's strongest model in the family. It is built for difficult work requiring advanced reasoning and stronger agentic capabilities, with notable improvements in areas including coding, biology and cybersecurity. The model introduces a new max reasoning effort for complex tasks. An additional ultra mode can use subagents to tackle demanding work, allowing the system to divide parts of a problem and bring the results together. Terra is positioned as a balanced option, combining strong intelligence with improved cost efficiency. Luna focuses on speed and lower-cost workloads, making the GPT-5.6 family adaptable to a wider range of applications. OpenAI says the new generation is designed to provide more useful intelligence from available computing resources while improving performance per dollar. The broader goal is to make advanced AI more capable for demanding professional, technical and agentic work. GPT-5.6 also reflects the growing shift from AI systems that primarily answer questions towards models capable of carrying out increasingly complex tasks and supporting more ambitious workflows. With Sol, Terra and Luna, OpenAI is expanding its frontier model strategy around different levels of reasoning power, efficiency and speed, giving users more flexibility in choosing the right intelligence for the work at hand.

@Matt_Pinner Are you thinking about the same industry that I'm thinking of? If you are also thinking about the same industry, then you are right.

🤔

@vxylily Thank Allah..

This will expose you

Alibaba's Qwen is disabling its humanlike interactive agents and user-created agent features as new Chinese regulations governing humanlike AI interaction services come into effect. Qwen announced that its humanlike interactive agents and user-created agent functions would be disabled from July 10, while broader Qwen agent functions and services would go offline on July 15. Following the shutdown, users will no longer be able to access related agent settings or previous conversations. The affected features allowed users to customise AI agents for particular tasks, skills and communication styles. Users could also create named assistants, tutors, role-playing characters or companions with specific personalities and tones. The changes coincide with China's new Interim Measures for the Administration of Artificial Intelligence Anthropomorphic Interaction Services, which take effect on July 15. The regulations cover AI services designed to simulate human personality traits, thinking patterns and communication styles while providing sustained emotional interaction. The rules do not generally cover customer-service bots, knowledge-based question-and-answer systems, workplace assistants, educational tools or scientific research applications, provided they do not involve sustained emotional interaction. Chinese authorities have cited concerns including privacy risks, physical and mental health, dependency, addiction and harmful content. At the same time, regulators continue to support the development of AI agents for practical and productivity-focused applications while introducing stronger standards around safety, identity, authorisation and traceability. The changes suggest a distinction in China's emerging AI regulatory approach between productivity-focused agents and humanlike companion agents designed to form sustained emotional or social relationships with users. ByteDance's Doubao is also disabling customised agent features as the new rules take effect, while Tencent reportedly removed a similar feature from its Yuanbao AI assistant in June.

Mistral AI has released Mistral OCR 4, a new document-parsing model that goes beyond simply extracting text by providing structured information about the content and layout of documents. The model introduces bounding boxes, block classification and inline confidence scores alongside extracted text. Each document block can be identified by its location and type, while confidence scores are provided at both page and word level. This gives downstream systems information about what a document says, where each element appears, what role it serves and how confident the model is in the extracted content. Mistral OCR 4 supports 170 languages across 10 language groups, including specialised and low-resource languages that many document-processing systems struggle to handle. The model accepts widely used enterprise formats including PDF, DOC, PPT and OpenDocument. As a compact and focused model, it can run within a single container and be fully self-hosted, allowing organisations with strict privacy or data-sovereignty requirements to keep sensitive documents within their own infrastructure. Mistral positions OCR 4 as an ingestion component for enterprise search, retrieval-augmented generation, domain-specific retrieval systems and agentic workflows. Its compact design also makes it suitable for cost-sensitive and high-volume document processing. The model is available through the Mistral OCR API or Document AI, giving organisations the choice between structured outputs for custom processing pipelines and more accessible document-processing applications.

Meta's LLaMA family has become one of the world's most widely adopted open-weight AI model ecosystems. First introduced in February 2023 as an alternative to proprietary models from companies such as OpenAI and Google, LLaMA has reportedly surpassed 650 million downloads. Unlike closed AI systems, Meta makes LLaMA model weights publicly accessible, allowing researchers, developers and businesses to download, modify and deploy the models without depending entirely on API access or recurring subscription fees. However, certain large-scale commercial uses remain subject to Meta's licensing conditions. LLaMA 4, released in April 2025, introduced a major architectural shift with a mixture-of-experts design and multimodal capabilities. The family includes LLaMA 4 Scout, with 17 billion active parameters across 16 experts and a context window of up to 10 million tokens, and LLaMA 4 Maverick, with 17 billion active parameters across 128 experts and 400 billion total parameters. Meta also announced LLaMA 4 Behemoth, a much larger model with 288 billion active parameters, from which Maverick was distilled. The LLaMA 4 family was designed to deliver competitive performance while maintaining lower inference costs than many proprietary alternatives. The earlier LLaMA 3 generation also played a major role in advancing open-weight AI. Released throughout 2024, the family included models ranging from 8 billion and 70 billion parameters to the frontier-scale 405 billion parameter model. LLaMA 3.1 expanded the context window to 128,000 tokens and improved multilingual capabilities, while LLaMA 3.2 introduced lightweight models designed for edge devices alongside multimodal vision capabilities. LLaMA 3.3 70B, released in December 2024, delivered performance comparable in some areas to the much larger 405B model while requiring significantly fewer resources to serve. Beyond standalone models, LLaMA technology has been integrated throughout Meta's ecosystem, helping power Meta AI experiences across WhatsApp, Instagram, Facebook and Messenger. Meta has also applied its AI technology to products including Ray-Ban Meta smart glasses and advertising tools. Meta's broader LLaMA ecosystem also includes safety tools such as LLaMA Guard for content moderation, Prompt Guard for protection against prompt injection attacks, and Code Shield for safer code generation. The company has also worked with major cloud providers to make LLaMA models more accessible for deployment. Through its continuing development of LLaMA, Meta is positioning open-weight AI as a major alternative to closed proprietary systems, supporting applications ranging from academic research and independent development to large-scale commercial deployments.

Google DeepMind has introduced Nano Banana 2 Lite, its fastest and most affordable image-generation model yet. Officially called Gemini 3.1 Flash Lite Image, the model is part of the Gemini 3.1 family and is designed to offer a balance between image quality, speed and cost. The new model is aimed particularly at rapid experimentation, idea exploration and fast prototyping. In its default low-thinking mode, Nano Banana 2 Lite can generate an image from a text prompt in about four seconds, compared with roughly 20 seconds for the standard Nano Banana model. That speed comes with some trade-offs. Google acknowledges that the Lite model may struggle more with very small text, can produce inaccurate information in infographics, and may show weaker consistency when generating the same characters or people across multiple images. For tasks requiring greater accuracy, better text rendering or photorealism, the standard Flash or Pro models may remain the better choice. However, user ratings on Arena.ai reportedly place Nano Banana 2 Lite surprisingly close to Google's more powerful image models in perceived output quality. Its lower cost could also make it particularly attractive to developers who need to generate images at scale or rapidly test multiple visual ideas. Google says Nano Banana 2 Lite averages around $0.034 per 1,000 images, while its API pricing is $0.25 per million input tokens and $1.50 per million output tokens, half the corresponding rates of Nano Banana 2 and significantly cheaper on output than Nano Banana Pro. All images created with Nano Banana 2 Lite include Google's SynthID watermark, designed to help identify AI-generated content even after editing. Nano Banana 2 Lite is available through Google AI Studio and the Gemini API. Users can also access it in Gemini by selecting the Flash-Lite option and requesting an image.

xAI has introduced Grok 4.5, describing it as its most capable model yet, with major improvements in coding, agentic tasks and knowledge work. The new model is designed to handle complex tasks that require sustained reasoning, tool use and multi-step execution. According to xAI, Grok 4.5 delivers stronger performance across demanding benchmarks and real-world workflows, reflecting the company's continued push toward more capable AI systems. Its improvements are particularly focused on areas where AI models increasingly operate as active assistants rather than simply answering questions. These include writing and debugging code, carrying out multi-step tasks, working with tools and handling complex knowledge-based assignments. Grok 4.5 represents xAI's latest step in advancing its flagship model family as competition among frontier AI developers continues to intensify.

Anthropic has introduced Claude Science, an AI workbench designed to bring scientific tools, computing resources and research workflows into a single environment. Scientific research often requires researchers to move between numerous databases, specialised file formats and tools such as PubMed, Jupyter, R and cluster terminals. Claude Science aims to bring these fragmented processes together, allowing scientists to analyse literature, conduct multi-step research, create detailed artifacts, and refine figures and manuscripts for publication. Every output includes an auditable history of how it was produced, helping researchers validate and reproduce their work. Claude Science can also generate rich scientific artifacts such as 3D protein structures, genome browser tracks and chemical structures, alongside the code used to create them. Researchers can request changes to figures using plain language, such as removing gridlines or changing an axis to a logarithmic scale, and the agent can modify its own code accordingly. The platform includes more than 60 curated skills and connectors across fields including genomics, single-cell research, proteomics, structural biology and cheminformatics. A coordinating agent can work with specialist agents, while a reviewer agent checks citations and calculations and flags or corrects errors. Claude Science is now available in beta for Claude Pro, Max, Team and Enterprise users.

OpenAI has introduced GPT-Live, a new generation of voice models designed to make AI conversations feel more natural and human-like. Built on a full-duplex architecture, GPT-Live can listen and speak simultaneously, reducing delays and allowing smoother, more natural conversations. It can acknowledge users with brief responses such as "mhmm," maintain the flow of discussion, and avoid interrupting while users are speaking. When a request requires deeper reasoning, GPT-Live hands the task to GPT-5.5 running in the background while continuing the conversation, creating a more seamless experience. During demonstrations, GPT-Live also performed real-time translation, including English to Hindi, while the speaker was still talking, eliminating the awkward pauses found in previous voice models. According to OpenAI, GPT-Live outperforms the earlier Advanced Voice Mode in both user preference and technical benchmarks. The rollout begins with GPT-Live-1 for Go, Plus and Pro users, while GPT-Live-1 Mini becomes the default voice model for free users. API access will follow.

Lenovo has introduced its AI Student Phone in China at an affordable price of ¥299, approximately $44 or ₹4,200. Designed to support education while reducing common smartphone distractions, the device comes without games, web browsers or social media apps. Instead, it focuses on AI-powered study assistance, educational resources and strong parental controls. The phone features a compact 1.83-inch (46 mm) touchscreen protected by Panda glass, handwriting input, a 1,850mAh battery and 4G connectivity. A dedicated physical AI button activates a voice assistant for quick educational support, while built-in learning resources include English vocabulary and maths formulas. Other features include QR code payments with spending limits, HD voice and video calls, and an SOS function. The device reflects growing interest in technology designed specifically around children's learning and safety. By removing entertainment-focused features while retaining essential communication and educational tools, Lenovo aims to help reduce distractions. Parents also receive greater oversight through features such as location tracking, geofencing and remote controls. Lenovo has introduced its AI Student Phone in China at an affordable price of ¥299, approximately $44 or ₹4,200. Designed to support education while reducing common smartphone distractions, the device comes without games, web browsers or social media apps. Instead, it focuses on AI-powered study assistance, educational resources and strong parental controls. The phone features a compact 1.83-inch (46 mm) touchscreen protected by Panda glass, handwriting input, a 1,850mAh battery and 4G connectivity. A dedicated physical AI button activates a voice assistant for quick educational support, while built-in learning resources include English vocabulary and maths formulas. Other features include QR code payments with spending limits, HD voice and video calls, and an SOS function. The device reflects growing interest in technology designed specifically around children's learning and safety. By removing entertainment-focused features while retaining essential communication and educational tools, Lenovo aims to help reduce distractions. Parents also receive greater oversight through features such as location tracking, geofencing and remote controls.

Apple has filed a federal lawsuit against OpenAI, OpenAI hardware chief Tang Tan, former Apple engineer Chang Liu, and io Products, alleging the misappropriation of Apple hardware trade secrets. The lawsuit follows months of reported tension and has also drawn attention to the strained relationship between Tan and his former boss at Apple, John Ternus. Apple alleges that Tan encouraged employees interviewing with OpenAI to bring physical Apple components for “show and tell.” OpenAI has denied wrongdoing and said it has no interest in competitors' trade secrets. The case highlights the growing competition for hardware talent as AI companies expand beyond software and into the development of physical devices. The claims made by Apple remain allegations and have not been established as fact in court.

@bushrants Like the Sol itself. With a surface temperature of 5,500°C.

haven't tried ChatGPT 5.6 sol yet , how is it???

GPT‑5.6 is now the preferred model in Microsoft 365 Copilot Today, OpenAI announced GPT‑5.6, which will become the new preferred model in Microsoft 365 Copilot—in Word, Excel, PowerPoint, Chat and Cowork. For Microsoft 365 customers, the update brings OpenAI's latest flagship model series into productivity tools people use every day, helping them create, analyze, and collaborate with more capable AI assistance across workstreams. GPT‑5.6 is OpenAI’s latest flagship model series, which delivers more useful work from every token, with stronger performance per dollar and on demand capability for the most complex tasks. With GPT‑5.6, Microsoft 365 users will be able to create higher-quality work products with less effort across the apps they already rely on: In Word, GPT‑5.6 can help people draft, edit, and refine documents with fewer rounds of prompting. In Excel, it can support deeper analysis while using tokens more efficiently, helping users move faster from data to insights. In PowerPoint, it can help turn early ideas into more polished, visually compelling presentations with less manual guidance. In Cowork, it can help users complete complex, cross-functional work and produce higher-quality outputs with less manual coordination. “We can’t wait for customers to see what GPT‑5.6 in Microsoft 365 will do, enabling them to work even more effectively with AI in the tools they use every day,” said Nitin Agrawal, President, Copilot & Agents Core, Microsoft. “Using Copilot powered by OpenAI’s latest model, customers will be able to produce more polished outputs in Word, Excel, PowerPoint, Cowork, and Copilot Chat, whether they are drafting documents, analyzing data, creating presentations, or collaborating across teams. We’re excited to continue building with OpenAI to bring more powerful AI experiences to people and organizations around the world.” “Microsoft 365 is where millions of people write, analyze, create, and collaborate every day,” said Nikunj Handa, Head of API Product, OpenAI. “By bringing GPT‑5.6 to Microsoft 365 Copilot through the OpenAI API, we're helping organizations get more useful work from every token, and more value from AI in the tools they already use.” In addition to serving the models natively, Microsoft will also access OpenAI models directly through the API to bring GPT‑5.6 to Microsoft 365 customers. Our partnership with Microsoft has always been about bringing the benefits of advanced AI to more individuals and organizations, and we’re excited to continue building on that shared commitment. — OpenAI