AF3 ➡️ Audio General Intelligence, Cognition acquired AI-coding startup Windsurf ......

NVIDIA Just Dropped Audio Flamingo 3: NVIDIA has released Audio Flamingo 3 (AF3), a fully open-source large audio-language model that brings human-like reasoning to speech, sound, and music. Unlike earlier models that transcribe or tag audio, AF3 can understand long-form audio (up to 10 minutes), hold multi-turn audio conversations, and generate voice responses—all while reasoning contextually across modalities. Trained on over 10 million audio-text pairs and fine-tuned with datasets like AudioSkills-XL and AF-Chat, it outperforms both open and closed models (like Gemini 2.5 Pro and Qwen2.5-Omni) on over 20 benchmarks, including ClothoAQA (91.1%) and LibriSpeech (1.57% WER).

Moonshot AI Releases Kimi K2-A Trillion-Parameter MoE Model Focused on Long Context, Code, Reasoning, and Agentic Behavior: Moonshot AI’s Kimi K2 is a trillion-parameter Mixture-of-Experts model designed for agentic AI workflows, featuring two variants—Base for fine-tuning and Instruct for immediate use. It supports long context windows of 128K tokens and is trained on 15.5 trillion tokens using the MuonClip optimizer for stable, large-scale training. Kimi K2 outperforms GPT-4 and Claude Sonnet 4 on coding and agentic benchmarks while costing about five times less per million tokens. Its open-source nature, native MCP support, and tool-use capabilities mark a shift from passive reasoning models to autonomous multi-step execution systems.

Microsoft's New Reasoning-Optimized Mini LLM: Microsoft has open-sourced Phi-4-mini-Flash-Reasoning, a 3.8B parameter model designed for efficient long-context reasoning. Built on the SambaY architecture with Gated Memory Units, it delivers up to 10× faster decoding and outperforms prior Phi models on benchmarks like Math500 and GPQA Diamond. Now available on Hugging Face.

Google AI Releases Vertex AI Memory Bank: Google Cloud has announced the public preview of Memory Bank, a new managed service within the Vertex AI Agent Engine. Memory Bank is designed to help you build highly personalized conversational agents that facilitate more natural, contextual, and continuous engagements.

Liquid AI has open-sourced LFM2: Their second-generation foundation models that deliver 2x faster inference than competitors while running directly on your device. The hybrid architecture combines convolution and attention mechanisms, enabling sophisticated AI on smartphones, laptops, and embedded systems without cloud dependency. Available in 350M, 700M, and 1.2B parameter versions under Apache 2.0 license, LFM2 outperforms larger models while using fewer resources—perfect for the growing edge AI market.

AWS debuted Amazon Bedrock AgentCore, which allows enterprises to securely deploy and operate AI agents at scale, plus new Marketplace offerings for agentic AI, and a $100 million investment boost

Google Cloud launched a new Conversational Agents Console and Agentspace, streamlining custom agent creation and integration using Gemini models and providing high-quality, natural human interaction for self-service experiences

CrowdStrike Security Integrations: CrowdStrike introduced new agentic AI-driven security workflow integrations for the AWS Marketplace, extending GenAI protection and streamlining operations for security teams

EPAM DIAL GenAI Platform Released: EPAM announced the availability of its DIAL GenAI platform on AWS Marketplace, offering new agentic AI integration capabilities across services

and many more

xAI is reportedly in talks to lease data center capacity in Saudi Arabia through partnerships with Humain and another data center operator, to scale its global training footprint

Following Grok chatbot controversies involving extremist and sexualized content, xAI began offering $180k–$440k roles to multimedia engineers in Palo Alto to build “flirty” real-time avatars, part of a recovery strategy

Cognition acquired AI-coding startup Windsurf following the departure of its leadership to Google and the collapse of a potential OpenAI buy. Integration is expected to influence pricing, interface, and tooling

MiniMax, a Chinese generative AI startup, filed confidentially for a Hong Kong IPO targeting a valuation north of $4 billion and aiming to raise roughly HK$4–5 billion (~$510–637 million) by year-end. Its multimodal models (text, audio, video, music) serve over 157 million individuals and 50k+ businesses globally

Perplexity CEO Aravind Srinivas publicly rejected potential acquisitions by major tech giants (Google, Meta, Apple), stressing a desire to stay independent. Rumors include a possible $500 million raise valuing the company at ~$14 billion

and many more…

SaneBox*: Smarter Email, Less Clutter
Automatically filters unimportant emails, letting you focus on what matters.
Learns your behavior to prioritize high-value messages and boost productivity.

AdCreative ai*: Generate Conversion-Optimized Ad Creatives
Creates ad visuals and texts tailored for performance across platforms.
Trained on real-world data to improve click-through and conversion rates.

Lovable – AI-powered app builder
Lovable allows you to build apps and websites entirely through natural language prompts—no manual coding required.

Replit – Code + Deploy in One Place
A collaborative cloud IDE with built-in AI support.
Write, test, and deploy applications without leaving your browser.

Gamma – Create Decks with a Prompt
An AI tool to turn ideas into polished presentations instantly.
Generate visually engaging slides by typing a single prompt.

Suno AI – Make Music Instantly
Generates original songs or instrumentals from simple prompts.
Perfect for creators, marketers, or anyone needing background music fast.

11x AI – 11x ai delivers autonomous “digital workers”
AI-driven agents like “Alice” (SDR) and “Julian” (phone agent)—designed to execute full-spectrum go-to-market activities around the clock.

_{Note: *We are affiliate partners with the above AI Tools and we get a referral fee when someone makes a purchase using the affiliate url link.}