Models And Releases - AI news 2025-07-23

Qwen3-Coder-480B-A35B-Instruct launches and it ‘might be the best coding model yet’

venturebeat.com • Jul 23, 2025

Developers can define custom tools and let Qwen3-Coder dynamically invoke them during conversation or code generation tasks....

TL;DR

Alibaba's open-source Qwen3-Coder model achieves leading performance in coding tasks, offering a strong alternative to proprietary models with its advanced capabilities and open-source Apache 2.0 license.

Key Takeaways:

Qwen3-Coder supports 256K token context lengths natively, with extrapolation up to 1 million tokens, enabling it to understand and manipulate entire repositories or lengthy documents in a single pass.
The model achieves leading performance among open models on several agentic evaluation suites, including SWE-bench Verified with 67.0% accuracy.
Qwen3-Coder is available under an open-source Apache 2.0 license, making it free for enterprises to use and modify in their commercial applications without any licensing fees.

Mixture-of-recursions delivers 2x faster inference—Here’s how to implement it

venturebeat.com • Jul 23, 2025

Mixture-of-Recursions (MoR) is a new AI architecture that promises to cut LLM inference costs and memory use without sacrificing performance....

TL;DR

Researchers at KAIST AI and Mila introduce Mixture-of-Recursions (MoR), a new Transformer architecture that makes large language models more memory- and compute-efficient.

Key Takeaways:

MoR delivers higher throughput and better model accuracy compared to vanilla transformers, with a 43.1% average few-shot accuracy and a 19% reduction in training time.
MoR's adaptive computation and parameter sharing design reduces memory traffic and improves throughput without needing complex post-training modifications.
The framework's scalability and efficiency gains make it a promising solution for enterprise adoption, particularly for tasks that require longer context windows and handling multiple samples simultaneously.

Google is bringing its AI-powered photo-to-video capability to more apps

www.theverge.com • Jul 23, 2025

Google is adding new AI-powered features to Google Photos and YouTube Shorts that allow users to transform their photos into videos. The tools are sim...

TL;DR

Google is rolling out AI-powered photo-to-video capabilities to Google Photos and YouTube Shorts, initially in the US.

Key Takeaways:

The new feature is available in Google Photos on Android and iOS devices in the US, with YouTube Shorts rollout starting in the US, Canada, Australia, and New Zealand.
Users can generate 6-second clips or select their own clip length in YouTube Shorts, with Google Photos restricted to making 6-second clips.
Google also releases a new Remix tool and AI Playground hub, with the former transforming photos into different styles and the latter providing generative AI effects for YouTube Shorts users.