AI Daily | Google Ads Advisor, Cohere Enterprise Model, and Stable Audio 3.0

Daily AI Pulse: Latest Breakthroughs in Google’s Marketing Assistant and Open-Source Models

New technology tools are being released every day. It’s truly exciting to watch these technologies mature step by step. Today’s AI Daily brings you the latest developments from major tech giants. We cover Google’s new AI agent for advertising, Cohere’s powerful open-source model designed specifically for enterprises, ByteDance’s lightweight multimodal powerhouse, and a brand-new gift for music creators from Stability AI. Let’s dive into these four noteworthy updates.

A New Partner for Ad Marketing? Introducing Google Ask Advisor

Ad placement can be a real headache. Marketers often have to switch back and forth between different data analysis platforms. Now, Google has launched Ask Advisor to solve this annoying problem. This is a cross-product AI agent that actually coordinates a whole team of expert agents behind the scenes, acting as a 24/7 collaborator and problem solver for marketers. It cleverly integrates resources from Google Ads, Google Analytics, and the Google Marketing Platform.

Just enter a natural language command like “find new customers for shampoo products.” The assistant will automatically pull product details from the Merchant Center and then directly create a new ad campaign. Sounds convenient, right? Users don’t need advanced data analysis skills. Ask Advisor explains which marketing strategies are working and provides specific suggestions for the next steps.

It can even proactively provide customized recommendations, saving teams significant time in trial and error. Many might wonder when this tool will become widely available. Currently, this feature is available in beta for English accounts, with more features rolling out in the coming months.

A Computing Powerhouse for Enterprises: Cohere Command A+

Next, let’s talk about Cohere’s release of Command A+. This is a highly attractive Mixture-of-Experts (MoE) architecture model. It’s designed for high-performance enterprise computing tasks and is the core driving force behind the progress of Cohere’s enterprise AI integrated workspace, “North.”

When it comes to enterprise applications, privacy control and hardware resources are always top concerns. Command A+ is released under the Apache 2.0 license. Most surprisingly, it can run smoothly on as few as two NVIDIA H100 GPUs, featuring an input context length of up to 128K and a maximum generation length of 64K. This significantly lowers the barrier for enterprises to deploy their own hardware. It has a total of 218 billion parameters, but only 25 billion active parameters are used for each activation.

To briefly explain, a Mixture-of-Experts architecture is like having specialized consultants for various fields within a large company. When a specific problem arises, the system only wakes up the corresponding consultant to handle it. Additionally, it employs speculative decoding technology optimized for the MoE architecture, boosting text and multimodal reasoning generation speeds by an additional 1.5 to 1.6 times. This not only significantly increases processing speed but also saves valuable computing costs. Its performance in multilingual support is also impressive, expanding from 23 languages to 48. Thanks to a new tokenizer, it particularly improves processing efficiency for Arabic, Korean, and Japanese, significantly reducing the number of tokens required for generation and inference costs.

A common question is where to download this powerful model. Developers can now directly go to Hugging Face or the Model Vault platform to obtain the weights and even experience its agent workflows.

A Small but Powerful Multimodal Star: ByteDance Lance Model

The third focus is the Lance multimodal model from ByteDance. You can also find complete test resources and architecture descriptions on the Lance page on Hugging Face.

When talking about multimodality, people usually think of those massive, beast-like models. But Lance takes the opposite approach. It has only 3 billion active parameters. This size is very compact, and more impressively, the development team completed the training process from scratch using at most fewer than 128 GPUs, demonstrating extremely high resource efficiency. Despite its size, it can handle multiple tasks like image generation, video understanding, and even video editing.

The development team trained this model completely from scratch, with its underlying foundation initialized based on the open-source Qwen2.5-VL, and built a “dual-expert architecture”—one specialized for understanding and the other for generation. They integrated all visual and text tasks under a single architecture. It’s like having an all-around assistant learn to draw, watch videos, and write simultaneously, with both cooperating within the same model without interfering with each other. Its performance in various open-source benchmarks is impressive. Many developers often wonder if such a small model can really handle video. As it turns out, Lance delivers satisfying results in multi-turn consistent editing and video generation several seconds long.

New Inspiration for Music Creators: Stable Audio 3.0

The final highlight belongs to artists and music lovers. Stability AI has officially unveiled Stable Audio 3.0. You can also view information on the full series of models in the Stable Audio 3 section.

The biggest feature of this version is its ability to generate audio up to six minutes long. The series is divided into four models to meet different needs: 3.0 Small SFX for 2-minute sound effects, 3.0 Small for 2-minute short tracks, 3.0 Medium for up to 6 minutes and 20 seconds, and 3.0 Large for the highest audio quality via API for enterprises. Music creation is a process full of colliding inspirations. Now, creators can fine-tune the model using LoRA technology to let the system learn their unique music style. This model uses a brand-new semantic-acoustic autoencoder, making generated tracks smoother and more natural. Even more amazing are the “inpainting” and “outpainting” features. This means you can replace a small segment of a track individually or continue extending from the end of a song without having to regenerate the entire piece.

Additionally, the 3.0 Small version can run offline directly on standard devices like laptops. This is very convenient for creators who are often on the go. Regarding copyright and commercialization, as long as a company’s annual revenue does not exceed $1 million, they can freely use and commercialize the generated music under a community license. This is undoubtedly an excellent creative aid for independent musicians.

The pace of technological development is always breathtaking. Watching these tools continue to evolve, future creation and work modes are sure to become even more interesting.

Q&A

Q1: What problem does Google’s Ask Advisor primarily solve, and how specifically can it help marketers? A: It primarily solves the pain point of marketers needing to switch back and forth between different platforms (like Google Ads and Google Analytics) to analyze data. Ask Advisor coordinates a “team of expert agents” behind the scenes. Users only need to enter natural language commands like “find new customers for shampoo products,” and it can automatically pull product details from the Merchant Center and create ad campaigns. It also analyzes data across platforms to explain marketing strategy effectiveness, and a beta version for English accounts is currently available.

Q2: Why is Cohere’s Command A+ particularly suitable for enterprise self-deployment? Are the hardware requirements high? A: The hardware requirements are very friendly! Command A+ is a Mixture-of-Experts (MoE) architecture model released under the Apache 2.0 license. Although it has a total of 218 billion parameters, it only wakes up 25 billion active parameters for each computation. Therefore, it can run smoothly on as few as two NVIDIA H100 GPUs. It also supports a context length of up to 128K and 48 languages (including optimized Japanese and Korean), significantly lowering the barrier and cost for enterprises to deploy high-end AI.

Q3: ByteDance’s Lance model is claimed to be “lightweight.” How small is it really, and can it actually handle video generation? A: Lance is very compact, with only 3 billion (3B) active parameters. The development team used fewer than 128 GPUs at most when training this model from scratch, showing extremely high resource utilization. Despite its small size, its unique “dual-expert architecture” (separating understanding and generation tasks to avoid interference) allows it to not only understand images and videos but also generate high-quality videos up to several seconds long (supporting up to 121 frames), and even perform complex multi-turn consistent edits.

Q4: What is the biggest draw of Stable Audio 3.0 for independent music creators? Can it be used offline? A: The highlights are generation length and fine-grained editing capabilities. The 3.0 Medium and Large versions can now generate complete songs up to 6 minutes and 20 seconds long. For editing, it supports “inpainting” and “outpainting,” meaning you can replace a small segment of a track or extend the end of a song without starting over. Even better, the 3.0 Small version supports running completely offline on standard laptops, and as long as annual revenue is under $1 million, the generated music can be commercialized under a community license.