DeepSeek unveiled DeepEP, an open-source EP communication library for MoE model training and inference. It supports efficient full-mesh communication, NVLink/RDMA tech, high-throughput/low-latency kernels, FP8 dispatch, and flexible GPU resource control, empowering developers globally to advance AI projects.

Alibaba launched QwQ-Max, a deep reasoning model in the Qwen series based on Qwen2.5-Max architecture. The preview version excels in logic, math, coding, and task management, surpassing DeepSeek R1 in tests. It supports tool integration for web searches and media generation, with plans for Apache 2.0 open-source release alongside smaller models like QwQ-32B.

DeepSeek released DeepEP, a communication library for Mixture of Experts (MoE) models. Optimized for GPU communication with support for NVLink and RDMA, it offers high-throughput, low-latency performance. DeepEP supports FP8/BF16 formats, hook-based computation methods, and is compatible with InfiniBand/RoCE networks. Key features include 153-158 GB/s intra-node and 43-47 GB/s inter-node bandwidth, with latencies as low as 163/318 microseconds. Requires Hopper GPUs, Python 3.8+, CUDA 12.3+, and

Perplexity AI, a generative AI search startup, is developing an AI-powered web browser named Comet, designed for "proxy search." This move aims to challenge dominant browsers like Google Chrome and Microsoft Edge by offering a more intuitive, AI-driven search experience. Comet may allow users to execute tasks via natural language prompts, redefining internet interaction. Despite controversies over content theft and lawsuits from publishers, Perplexity AI continues expanding its offerings, includ

Salesforce and Google Cloud signed a $2.5 billion, seven-year partnership to integrate AI, CRM, and big data capabilities. Salesforce customers can now run Agentforce AI agents on Google Cloud's infrastructure, leveraging advanced analytics tools. This collaboration aims to compete with Microsoft for enterprise clients, offering seamless integration and improved AI functionalities. Key benefits include using Google Docs with Salesforce data and enhancing Agentforce's multimodal support. The move

MongoDB acquires Voyage AI, a startup specializing in embedded AI models, to enhance MongoDB Atlas capabilities. Voyage AI's flagship algorithm outperforms competitors in embedding quality and offers specialized models for various data types. The integration will allow developers to store standard data and embeddings within Atlas, simplifying workflows and enabling multi-modal data support for industries like law and finance.

Elon Musk's xAI launches Grok 3, powering Grok's chatbot and enhancing X platform. Despite competition from ChatGPT, Grok 3 saw a 10x mobile download increase and 260% U.S. daily active users growth post-launch. Expansion into new markets and web traffic surges indicate strong initial adoption, but sustaining user engagement remains critical.

Apple plans to invest over $50 billion in the US over four years, including building an AI server plant in Texas and creating 20,000 research jobs. The company will expand its Advanced Manufacturing Fund to $10 billion, support TSMC's Arizona plant, and launch an Apple Manufacturing Academy in Detroit. This initiative aims to boost domestic innovation, enhance AI capabilities, and address political pressures on manufacturing.

Anthropic launched Claude 3.7 Sonnet, the first "hybrid reasoning model" that combines instant responses with step-by-step thinking. It excels in coding tasks and integrates Claude Code, a tool for code editing and GitHub integration. Pricing remains at $3 per million input tokens and $15 per million output tokens. The model offers flexible thinking token control and improved security, reducing false rejections by 45%. Extended thinking is available across paid plans and platforms like Anthropic

A recent OpenAI study reveals that despite rapid advancements, top AI models still fall short of human programmers' capabilities. Using a new benchmark, SWE-Lancer, researchers tested OpenAI's o1 and GPT-4o, and Anthropic's Claude 3.5 Sonnet on real-world coding tasks. While fast and capable of addressing superficial issues, the models struggle with complex errors and lack contextual understanding, providing incomplete solutions. The study concludes that LLMs, though efficient for specific tasks

PixVerse V4 introduces automatic sound effect generation, a voiceover function, and video re-styling options to enhance creativity. It features realistic AI-generated content, customizable styles, and ultra-fast video creation in just 5 seconds, setting a new global standard for AI-driven video tools. Available now for all users.

StepStellar's latest video generation model, Step-Video-T2V, is now integrated into LiblibAI, making it the exclusive AIGC partner for its open-source release. Users can access the model via LiblibAI's platform, which has over 20 million creators and generates millions of interactions daily. This collaboration aims to democratize AI technology and drive innovation in multimodal content creation.

The Moonlight-16B-A3B model, developed by Kimi's team using Moonshot AI, features 160 billion parameters with only 30 billion active for efficiency. Trained on 5.7 trillion tokens, it excels in benchmarks like MMLU and HumanEval. Using an optimized Muon optimizer, it achieves faster, stable training and reduces computational costs, offering advanced NLP capabilities.

The DeepSeek-R1 model on Hugging Face has garnered over 10,000 likes, making it one of the most popular models among 1.5 million on the platform. Hugging Face launched the Open-R1 project to enhance DeepSeek R1's open-source ecosystem by disclosing previously undisclosed datasets and code, promoting research and innovation. The planned synthetic dataset will span multiple disciplines, boosting Hugging Face's influence in the AI community.

DeepSeek launches Open Source Week with FlashMLA, an optimized multi-head linear attention decoding kernel for NVIDIA Hopper GPUs. Inspired by flashAttention and Cutlass, it enhances memory efficiency with BF16 format and a paging KV cache. Supporting CUDA 12.3+ and PyTorch 2.0+, FlashMLA achieves up to 3000 GB/s memory bandwidth and 580 TFLOPS on H800 GPUs. Its block scheduling and parallel computation optimize performance for large-scale data tasks. Easy installation via python setup.py instal

OpenAI is reportedly shifting its cloud strategy, reducing reliance on Microsoft Azure by gradually transitioning to Stargate, a new AI cloud infrastructure project. Backed by SoftBank, Stargate aims to invest $500 billion in US AI infrastructure over four years. By 2030, OpenAI expects three-quarters of its computing capacity to come from Stargate, potentially reshaping the AI cloud market. Recent changes in OpenAI's partnership with Microsoft allow it to explore other providers like Oracle whi

Apple's iOS 18.4 Beta 1 introduces AI-powered Priority Notifications, using on-device intelligence to highlight critical alerts at the top of the lock screen. This feature organizes notifications based on importance rather than time, focusing on urgent messages or reminders. Disabled by default, users can enable it via Settings > Notifications. The update also supports ten new languages and aligns with Apple's broader AI-driven initiatives like the Intelligence Suite. However, Siri's AI upgrades

Microsoft's Majorana 1 chip uses topological qubits for scalable, reliable quantum computing, potentially solving problems beyond classical capabilities. The compact design integrates error correction and advanced materials like indium arsenide and aluminum, enabling stability and precision. This breakthrough could revolutionize healthcare, sustainability, and logistics.

Figure AI's HELIX integrates vision, language, and action into a unified neural network, enabling humanoid robots to perform complex tasks with minimal programming. Trained on diverse datasets, it offers high adaptability, energy efficiency, and collaborative capabilities for industrial and home use. Future development will focus on enhancing real-world autonomy and voice integration.

Google revealed pricing for its AI video model Veo 2: $0.5 per second ($30/minute or $1,800/hour). For context, Jon Barron compared it to "Avengers: Endgame's" $32,000/second budget. Veo 2 is designed for clips over two minutes, unlike full films. OpenAI's Sora costs $200/month via ChatGPT Pro.