Plus, Udio 1.5 Unveils Audio-to-Audio Capabilities and Mixable Stems
Introduction
This is the inaugural edition of NexaQuanta\’s AI Insights, where we bring you the latest advancements in AI technology, tools, models, and research. In this edition, we cover:
- The new licensing and cost efficiency of DeepSeek\’s V2 model
- Stable Video\’s leap into four-dimensional video
- A closer look at how Runway trains its video models
- Microsoft’s latest updates to the Phi-3 family, making fine-tuning and deployment more accessible
But first, let’s dive into:
Mistral’s Powerful New Model
Mistral AI has launched Mistral Large 2, an advanced language model with 123 billion parameters and a 128,000-token context window.
This model supports a wide array of languages and over 80 coding languages, and it is designed to redefine performance and cost efficiency.
In its pre-trained version, Mistral Large 2 achieves 84% accuracy on MMLU and outperforms many of its predecessors, including Claude 3 Sonnet and GPT-4.
As the company shifts focus to its most advanced models, Mistral Large 2 will be available on the Mistral platform and through major cloud providers. It will have flexible licensing for research, non-commercial, and commercial use.
Udio 1.5 Enhances Musical Creativity
The latest update from Udio introduces stem downloads, enabling users to separate tracks into distinct elements such as vocals, bass, and drums for more refined mixing and remixing.
Additionally, the new audio-to-audio feature allows users to upload and transform their own tracks using AI, while key control options offer more precise harmonic adjustments.
These tools give musicians and producers greater control over AI-generated music, opening up new creative avenues for novices and professionals.
DeepSeek-V2 Licensed for Broader Use
DeepSeek has redefined the licensing of its DeepSeek-V2, a 236 billion parameter mixture-of-experts language model.
Thanks to innovative attention and feed-forward network architectures, this model excels in performance while reducing training costs by 42.5% compared to its predecessor.
With outstanding results across tasks in English, Chinese, coding, and math, DeepSeek-V2 is now available under a custom license that supports commercial use.
The code repository is also released under the MIT License. API access offers millions of free tokens to new users and a cost-effective pay-as-you-go model for ongoing use.
Stable Video 4D: A New Era in Generative Video Research
Stability AI has launched Stable Video 4D, an innovative model that generates eight distinct videos from a single object video by specifying 3D camera poses.
Producing 5-frame videos across eight views in just 40 seconds, this model enhances consistency across spatial and temporal dimensions, pushing the boundaries of generative video technology.
Currently available on Hugging Face, Stable Video 4D is in the research phase, with further improvements underway.
Runway’s Training Practices Come Under Scrutiny
Documents suggest that Runway may have used thousands of YouTube videos and pirated content to train its Gen-3 video generation model.
An internal spreadsheet obtained by 404 Media reveals that the company sourced videos without consent from well-known YouTube channels, influencers, and media outlets.
This revelation sheds light on Runway’s training methods while raising ethical concerns about data collection practices, particularly in light of Google’s stance against such scraping.
Microsoft Streamlines AI Development with New Phi-3 Updates
Microsoft has introduced significant enhancements to its Phi-3 family of language models, including serverless fine-tuning for Phi-3-mini and Phi-3-medium and the release of Phi-3-small via a serverless endpoint.
These updates make AI development more accessible, enabling developers to quickly build applications without the need to manage infrastructure.
The improvements to Phi-3-mini also enhance its performance in areas like instruction-following and structured output, making it a versatile tool for cloud and edge scenarios.
Conclusion
As this is NexaQuanta\’s first newsletter, we’re excited to bring the latest and most relevant AI developments to your inbox. In future editions, we’ll continue to deliver insights and analyses on new trends and innovations shaping the world of AI.
Stay tuned for more updates and stories that matter in AI, including breakthroughs in models like OpenAI’s GPT4-o mini, regulatory changes affecting AI in the EU, the latest on AI chip investments, and emerging technologies like VASA-1’s generative talking-head video system.
Subscribe to NexaQuanta\’s Weekly Newsletter
Your Guide to AI News, Latest Tools & Research
[contact-form-7 id=\”db89d1f\” title=\”Newsletter\”]
