👾 On-Device AI, Agentic Video Editors & Claude Vibe-Coding

Google’s lightweight Gemma 3n model leads benchmarks with < 3GB of memory. HeyGen’s agent video team, new models from China, and Claude Code in your IDE.

Author

Tobias Wupperfeld
July 01, 2025

Welcome to The Agent Roundup

This week, on-device intelligence leaps with Google’s lightweight Gemman 3n, while HeyGen’s creative agents now script, cast, and cut full videos with just a few inputs.

Tencent and Alibaba drop fresh models, Claude Code levels up your vibe-coding game, and even Wimbledon swaps line judges for AI.

Oh, and Nvidia? It just became the most valuable company on planet Earth.

Let’s dive in. 🌀

This week’s topics:

  • High performance on-device AI with Gemma 3n

  • HeyGen’s video agent creates video content from simple text

  • China answers with Tencent Hunyuan-A13B and Alibaba Qwen-VLo

  • Next level vibe coding with Claude Code

  • Plus AI investments, trending AI tools, community highlights, and more

AI Agent News Roundup

💥 Breakthroughs

Gemma 3n Improves On-Device AI

Source: Google

Google’s new open-source Gemma 3n LLM delivers enhanced quality, becoming the first <10B model to pass 1300 on LMArena.

It’s multimodal by design, supporting image, audio & video inputs.

Built for on-device use cases. Architectural innovations allow it to run with as little as 2GB - 3GB of memory.

HeyGen Video Agent Cuts Videos for You

Source: HeyGen

HeyGen introduced a Creative Operating System that creates video content with scripts, actors, and edits.

All it needs are simple texts, product images, or videos.

An entire team of agents analyzes the inputs and creates realistic ads, TikToks, short films, product demos, and more.

New Models from Tencent and Alibaba

Windsurf SWE-1 Coding Models

Source: Windsurf

Tencent released a new open-source hybrid reasoning model, Hunyuan-A13B.

It delivers high-performance results while achieving optimal resource efficiency.

Alibaba released a preview of Qwen-VLo. It’s capable of image generation and offers advanced multimodal understanding.

📈 Investments

🇺🇸 Nvidia has just become the world’s most valuable public company, with its $3.8T market cap, surpassing both Microsoft and Apple, providing 80% of the world’s AI chips.

🇺🇸 Salesforce CEO Marc Benioff stated that AI is handling 30-50% of the company’s workload, at about 93% accuracy.

🇨🇳 Xiaomi released its first AI-powered glasses, featuring an AI assistant with voice commands, a 12MP camera, and 2x the battery life of the Ray-Ban Meta glasses.

🇨🇳 Zhipu AI launched AutoGLM Rumination, a free AI agent capable of autonomous research, internet searches, itinerary planning, and report writing. It competes with leading US AI systems.

🇬🇧 Wimbledon replaces human line judges with AI-powered ball-tracking technology for the 2025 tournament, delivering more consistently accurate calls.

Inventory Software Made Easy—Now $499 Off

Looking for inventory software that’s actually easy to use?

inFlow helps you manage inventory, orders, and shipping—without the hassle.

It includes built-in barcode scanning and connects with Shopify, Amazon, QuickBooks, and more.

Try it free and save $499 with code EASY499—for a limited time.

🚀 Compare plans on our pricing page

Tool Spotlight

👾 Claude Code in VSCode: Replacing Cursor and Windsurf?

Claude Code is a coding agent you can control right from the terminal and easily plug into VSCode or Cursor.

Install NodeJS 18+, then run npm install -g @anthropic-ai/claude-code in your terminal. Then type claude and follow the instructions. When you run Claude from the terminal within your IDE, it auto-installs the extension.

What Are The Advantages?

  • It has better context over your entire codebase and is native to Claude

  • You can set up custom system prompts or commands, and agents working in parallel in the background

  • Better at following existing code patterns and handling complex debugging and project management tasks

What are the Downsides?

  • Only works with Claude models, which are more on the expensive side

  • Less visual integration, more command-driven

  • It’s slower, takes more time for reasoning

Level up your vibe coding by combining both tools: using Claude Code for internal, complex tasks like debugging and Cursor for quick, less complex tasks.

You can run Claude Code in multiple terminals at the same time. This is great for parallel feature development or design iterations. Then you pick the best version or cherry-pick.

Here is a list of common workflows and commands to try out.

Anthropic also released the Claude Code SDK. It lets you integrate Claude’s features into other applications.

🎛️ Gemini CLI: An open-source terminal agent similar to Claude Code that is powered by Gemini 2.5 Pro. It offers high free usage limits.

🧰 Claude Artifacts workspaces: A platform allowing any user to create, host, and share AI-powered apps, vibe-coded in Claude.

📸 Higgsfield Soul: A high-aesthetic photo model with advanced realism and 50+ presets for easy style optimization.

🎧 Eleven v3: A text-to-speech model with advanced realism and expressions. It supports 70+ languages with accurate accents.

Community Highlights

More Resources

Blog: In-depth articles on AI workflows and practical strategies for growth
AI Tool Collection: Discover and compare validated AI solutions
Consultancy: Explore AI potential or make your team AI-fit
Agency: AI implementation services to scale your business

See you next time!

Tobias from The Agent Roundup