What's New in AI This Week: Google, OpenAI Chat Memory, Llama 4 and Open Source Models?

TL;DR

Google led this week’s AI releases with Firebase Studio, the sixth-generation Ironwood TPU, Gemini 2.5 Flash and public VEO 2 access, while Meta launched Llama 4 Scout and Maverick with 16 billion active parameters each. Open-source image models also arrived for generation and stylization, including HiDream AI and a Flux.1-based model. Read on for specific capabilities, hardware requirements and early limitations.

Transcript

This week has been absolutely colossal with AI news. I mean, so many different releases from big names in the tech space, little open- source tidbits, cool demos, you name it, we got it today. These are the kinds of weeks that I like to see in the AI world. So, let's just jump right in. First up, quick little recap. Llama dropped from Meta. I did a... Read More

Key Insights

• Meta's Llama 4 released two models, Scout and Maverick, both with 16 billion active parameters, but Scout uses 16 experts while Maverick uses 128 experts. Neither runs on consumer-grade hardware, as both are aimed at corporations and businesses.
• HiDream AI is a brand new MIT-licensed image generation model with full, dev, and fast variants. Its benchmarks reportedly beat Dolly 3, SDXL, and Flux, though not by much, and it requires significant VRAM to run.
• Google's Ironwood is its sixth-generation TPU built for AI inferencing, offering 192GB of RAM per chip and 4.5 times faster data access, positioned as a cheaper alternative to Nvidia GPUs.
• Firebase Studio is Google's AI vibe-coding platform that automates typical coding processes, but it runs on a weaker Gemini model rather than 2.5 Pro, leading to less than satisfactory results in early testing.
• Google's VEO 2 is now public and usable directly in Gemini, though image uploading is not supported there. The API adds inpainting, outpainting, camera presets like panning, and first and last frame control.
• Gen 4 Turbo is now available at five times faster and half the cost of the original Gen 4. It trades some quality and prompt coherence for rapid ideation and struggles with large non-human motion and complex prompts.
• A new research paper, one-minute video generation with test-time training, produces coherent full one-minute Tom and Jerry cartoons with characters interacting logically, and could potentially extend to five or ten minutes.
• 11 Labs launched a new MCP server that gives Claude and Cursor access to its entire AI audio platform through simple text prompts, enabling use cases like spinning up voice agents to perform outbound calls.
• "Both have 16 billion active parameters, but Scout only has 16 experts, while Maverick comprises of 128 experts." (0:29)
• "Neither models, even the smallest one, Scout, run on consumer grade hardware." (0:38)
• "192 GB of RAM per chip and it's got 4.5 faster data access." (5:21)
• "In the API, they have inpainting and outpainting features, which is pretty awesome." (5:57)

Questions & Answers

Q: What are the differences between Llama 4 Scout and Maverick?

Llama 4 Scout and Maverick both have 16 billion active parameters. Scout has 16 experts, while Maverick has 128 experts, and neither model runs on consumer-grade hardware.

Q: Who are Llama 4 Scout and Maverick designed for?

Both Llama 4 models are intended for corporations and businesses rather than consumer-grade systems. They remain open source, although the transcript says their license is not the greatest.

Q: What is HiDream AI and where can you test it?

HiDream AI is an open-source image generation model with full, dev and fast quality variants. It is MIT licensed, while its Llama 3-based text encoder retains the Llama 3 license, and all three variants can be tested free on Hugging Face.

Q: How much hardware does HiDream AI require?

HiDream AI requires a significant amount of VRAM to run. The Hugging Face demonstrations use a quantized version, so they do not fully represent the largest version of the model.

Q: How well does Firebase Studio work?

Firebase Studio is an AI vibe-coding platform that automates typical coding processes using Gemini. It uses a weaker model instead of Gemini 2.5 Pro, and early examples and comments reported unsuccessful app generation and launch environment issues.

Q: What is Google’s Ironwood TPU?

Ironwood is Google’s sixth-generation TPU for AI inferencing. It has 192 GB of RAM per chip and 4.5 faster data access, and the transcript presents it as a potentially cheaper alternative to Nvidia GPUs.

Q: What can you do with VEO 2 now that it is public?

VEO 2 can be used directly inside Gemini, although Gemini does not support image uploading for it. Its API includes inpainting, outpainting and camera presets such as panning to the right.

Q: How does the Flux.1-based image stylization model compare with GPT4 Omni?

The open-source stylization model is built on Flux.1 and is intended to provide controls similar to GPT4 Omni. The transcript says it is less effective at preserving an uploaded character and transferring style, but it is faster, Apache 2.0 licensed and available free on Hugging Face.

Summary & Key Takeaways

• This was a colossal week for AI news across major companies and open source. Meta's Llama 4 Scout and Maverick launched with 16 billion active parameters each but need corporate hardware, drawing a meh reception. Ply had already jailbroken Llama 4 shortly after release.

• Google dominated the week, launching Firebase Studio for AI coding on a weaker Gemini model, the Ironwood sixth-gen inferencing TPU with 192GB RAM per chip, updated Imagen generation, Gemini 2.5 Flash, and public VEO 2 access through Gemini and a feature-rich API.

• AI video generation surged with Gen 4 Turbo running five times faster at half the cost, Higgsfield AI's new combinable camera motion controls, and LTX Studio adding consistent custom AI actors. Open image models HiDream and a Flux-based stylizer arrived free, and 11 Labs shipped an MCP server.