The AI landscape is moving fast with three major announcements this week:

• Google introduced EmbeddingGemma, a lightweight 308M-parameter multilingual embedding model optimized for on-device AI. With support for Matryoshka Representation Learning, it delivers top-tier performance in semantic search and retrieval while staying efficient for mobile and offline use.

• OpenAI unveiled its upcoming AI-powered hiring platform to rival LinkedIn, aiming to connect AI-fluent talent with businesses and governments. Paired with OpenAI Certifications via the OpenAI Academy, the initiative seeks to train millions in AI by 2030, ensuring workforce readiness for the future.

• Moonshot AI launched Kimi K2-0905, a trillion-parameter mixture-of-experts model with a 256K-token context window, advanced coding agent capabilities, and state-of-the-art results on programming and reasoning benchmarks.

✨ Let’s explore how these breakthroughs are shaping the next wave of AI innovation.

Google’s Efficient On-Device Multilingual Embedding Model

Google’s EmbeddingGemma is a lightweight text embedding model built for on-device AI. With 308 million parameters, it’s small enough to run smoothly on mobile and offline setups while still delivering strong performance. It creates 768-dimensional embeddings but can also shrink them to smaller sizes (512, 256, or 128) using Matryoshka Representation Learning, making it faster and more storage-friendly. Despite its compact size, it ranks among the top models under 500M parameters on the Massive Text Embedding Benchmark (MTEB), especially for multilingual search and retrieval. Plus, it works easily with Hugging Face, LangChain, and LlamaIndex, enabling private and scalable offline AI pipelines.

OpenAI Announces AI-Powered Hiring Platform to Compete with LinkedIn

OpenAI believes AI has the potential to unlock unprecedented economic opportunities by helping companies operate more efficiently and by empowering individuals to turn ideas into income and new types of jobs. While AI will cause disruption and change how work is done, OpenAI is committed to helping people become fluent in AI and connect with companies seeking these skills to foster economic inclusion. They are building the OpenAI Jobs Platform to match AI-savvy candidates with employers, including local businesses and governments, and launching OpenAI Certifications via the OpenAI Academy to train and certify millions in AI fluency by 2030. This initiative, supported by partners like Walmart, aims to ensure AI skills translate into real job opportunities and economic growth, focusing on the needs of both workers and employers. OpenAI's approach emphasizes practical, employer-grounded training and access to AI tools to shape a future where everyone can benefit from AI-powered innovation and jobs.

Google’s NotebookLM Adds New Audio Overview Formats

Kimi K2-0905 is the latest update to Moonshot AI's state-of-the-art mixture-of-experts (MoE) language model with 1 trillion total parameters and 32 billion activated parameters. This release significantly extends the context window from 128K to 256K tokens, enabling the model to handle much larger texts, codebases, or multi-turn conversations with ease. It enhances agentic coding capabilities, especially for front-end development and tool-calling, improving reliability and integration with various coding agents. The model also introduces improved factual grounding and reduces hallucinations in programming and structured output tasks. Kimi K2-0905 employs a custom MuonClip optimizer for training stability and delivers top-notch performance on benchmarks like LiveCodeBench and AIME 2025. Available via GroqCloud and other platforms, it supports expansive, complex AI-driven workflows in coding, research, and problem-solving.

Hand Picked Video

In this video, we’ll look at how Google’s Gemini Nano Banana model edits complex images, replacing objects, improving vibes, restoring photos, and even making creative transformations.

Top AI Products from this week

• CapCut AI Suite – Introducing CapCut AI Editing - Your smart editing partner for camera + prompt-led creation. From auto edits to avatars, smart cuts, and asset generation, it’s AI editing the CapCut way: simple, powerful, smart.

• 100 Vibe Coding – maestro is our bold step in building an ethical music generation model. Trained on 3.5M licensed tracks, this foundation model is capable of crafting high quality music with text prompts. Your personal AI composer is now more creative and inspirational.

• Clips by xdge.ai – Easily record and share clips to eliminate meetings. An AI note-taker can record, autonomously attend, and create speaker-tagged transcripts, summaries, action items, and coaching insights. Notes become part of a replayable collection of workplace knowledge.

• TextJam- TextJam is a multi-player AI editor with novel twists on how humans co-write with AI (and each other). From typing in pen and pencil, to inline prompting, to multi-touch gestures that resize text, TextJam is a bold new take on what a word processor can be.

• Stax – Your AI posse for researching, planning & building software. Repo-aware, multi-agent, and local-first — Shotgun creates decision-ready specs and shareable artifacts that truly capture your intents and context, so AI tools finally build what you mean.

• Brick Coder AI – A weekly meta-newsletter that cuts through AI noise. Every Monday, it delivers the most important research, models, tools, deployments, industry moves, and governance updates. Curated, fact-checked, and linked to primary sources.

• bestie - Test your AI agents with AI users across different personas and scenarios. Catch issues before they hit your real users.

This week in AI

• NVIDIA AI Blueprint for 3D - NVIDIA launches an AI Blueprint for 3D object generation with Microsoft TRELLIS NIM, enabling artists to create and prototype 3D scenes 20% faster from simple prompts.

• Create Your Own Android Bot With AI - Google launches Androidify, letting you create custom Android bots from selfies or prompts, with Gemini 2.5 Flash, Imagen, and Veo 3 for animations.

• OpenAI’s First AI Chip - OpenAI will launch its first in-house AI chip in 2026 with Broadcom, aiming to cut reliance on Nvidia and power its own AI systems, reports the Financial Times.

• DeepSeek’s Next AI Agent - China’s DeepSeek is preparing a new agentic AI model for late 2025, designed to perform multi-step tasks, learn from actions, and rival OpenAI in advanced AI agents.

Paper of The Day

A hybrid flight trajectory planning approach that integrates Reinforcement Learning (RL) with traditional search-based path planners. The RL agent pre-computes near-optimal coarse routes considering location and atmospheric data, which then constrain the search space of a detailed path planner. This hybrid method achieves up to a 50% reduction in computation time with negligible loss in fuel efficiency compared to conventional planning. The model is designed for rapid online re-routing, crucial in emergency flight diversions, and shows promise for accommodating weather uncertainty in future work. Experiments with Airbus aircraft models validate the approach's effectiveness in speeding up route optimization while maintaining solution quality.

To read the whole paper, go to here.