Elon Musk’s xAI Unveils Grok-3: Is This the World’s Most Powerful AI Yet?

Elon Musk’s xAI has unveiled Grok 3, a cutting-edge AI model that surpasses competitors like GPT-4o, Claude 3.5, and Gemini 2 Pro in mathematics, coding, and science. It was trained using 200,000 Nvidia H100 GPUs in a custom-built Memphis data center, which xAI constructed in just over 200 days.

Key Highlights:

• Performance: Outperforms top AI models on benchmarks like AIME 2025 and GPQA.

• Capabilities: Conducts deep internet searches (DeepSearch), predicts trends, and generates real-time game content.

• Reasoning Models: Features a "Think" mode and a high-powered "Big Brain" mode for complex problem-solving.

• Infrastructure: xAI is now building a 1.2 GW supercluster—the largest in the world.

• Commercial Access: Initially available via X Premium+ ($50/month) and SuperGrok ($30/month) for advanced features.

• Open-Source Plans: Grok 2 will be open-sourced once Grok 3 stabilizes.

Link to video 

Microsoft OmniParser Help See, Understand, and Control Your Screen

Microsoft has released OmniParser V2, an advanced screen parsing tool that converts UI screenshots into structured data, enhancing LLM-based UI agents.

Key Improvements in V2:

• Larger and cleaner dataset for icon captioning and grounding.

• 60% reduction in latency (0.6s/frame on A100, 0.8s on a single 4090 GPU).

• Higher accuracy (39.6 on ScreenSpot Pro, a major improvement from GPT-4o’s 0.8).

• OmniTool integration, allowing control of Windows 11 VMs with OmniParser and vision models (e.g., OpenAI, DeepSeek, Qwen, Anthropic).

Why It Matters:

General-purpose LLMs struggle with GUI automation due to challenges in identifying and interacting with UI elements. OmniParser ‘tokenizes’ UI screenshots, making them interpretable for LLMs, enabling better next-action predictions. With enhanced accuracy and efficiency, V2 is a significant upgrade for GUI automation tasks.

Model: https://huggingface.co/microsoft/OmniParser-v2.0
Demo: https://huggingface.co/spaces/microsoft/OmniParser-v2

Perplexity Launches Deep Research AI Agent to help you do research

Perplexity has launched Deep Research, a new AI-powered research agent that conducts expert-level analysis and delivers comprehensive reports in minutes. It is free for all users, with Pro subscribers ($20/month) getting unlimited queries and faster responses. The tool autonomously searches, reads, and synthesizes information, performing research that would take humans hours. It excels in finance, marketing, technology, and personal consulting (e.g., health and travel planning). Users can export reports to PDFs or share via Perplexity Pages. Deep Research is available on the web now, with mobile and Mac support coming soon.

Try it at perplexity.ai by selecting “Deep Research” in the mode selector.

Link to article

Launching Open Deep Researcher: Your New AI-Powered Research Assistant

I’m thrilled to introduce my latest project, a free and open-source alternative to hashtag#OpenAI Deep Research and Perplexity Deep Research. An autonomous agent that synthesizes vast amounts of online information to complete multi-step research tasks for you!

What It Does:

 ✅ Conducts in-depth research & analysis autonomously
 ✅ Performs dozens of searches & reads hundreds of sources
 ✅ Uses reasoning to synthesize complex information
 ✅ Delivers comprehensive reports on expert-level topics

Why This Matters:

Deep Research can save you hours by doing the heavy lifting analyzing information, summarizing insights, and providing well-structured reports so you can focus on decision-making instead of data hunting.

Mistral Saba 24B for Middle Eastern and South Asian Languages

Mistral has released Saba 24B, a lightweight language model designed for Arabic and South Asian languages like Tamil and Malayalam. Despite its compact size (24 billion parameters), it outperforms models five times larger while running efficiently on a single GPU.

Key Highlights:

• Superior Arabic Performance: Outperforms 70B+ models in Arabic benchmarks.

• Cost-Efficient & Fast: Processes over 150 tokens per second.

• Cultural Awareness: Delivers more relevant responses by understanding regional nuances.

• Flexible Deployment: Available via API and on-premise for industries like finance and healthcare.

Strategic Impact:

Mistral is positioning itself in the Middle Eastern market, potentially attracting regional investors. The model strengthens its reputation as an international AI alternative to U.S. and Chinese competitors.

Use Cases:

• Conversational AI: Powers Arabic virtual assistants.

• Domain-Specific AI: Fine-tuned for industries like energy and healthcare.

• Cultural Content Generation: Creates authentic, localized content.

Mistral Saba marks a step towards AI models that are not just multilingual but truly native to specific regions.