5/26/2025 AI Express | AI Field Update: Multimodal Reasoning, Intelligent Assistants & Robotics Update

BAGEL is an end-to-end graph comprehension and generation multimodal reasoning model open-sourced by ByteDance. It is capable of realizing the native graph generation capabilities of GPT-4o and Gemini 2.0, breaking the monopoly of closed source models.

BAGEL not only understands and generates images, but also performs multiple rounds of dialog editing with a wide range of free-form graphic generation and comprehension capabilities, including image editing, inference, combinatorial modeling, world modeling, and more.

The model is intended to be a "unified" big model, able to understand and generate text, images, videos, etc. simultaneously, without the need for multiple models to be responsible for different tasks and complex collaboration.

BAGEL hasinference chain mechanism(Reasoning Chain), i.e., before generating images or editing content, the model will first "think for itself" what to do in words. For example, when a user asks "turn the cat in the picture blue", BAGEL will first organize the following steps internally:

1. Confirmation that there's a cat in the image.
2. Identify the exact location of the cat
3. Determine the details to be retained
4. Consider how to change the color without destroying the background

This "think before you act" approach allows BAGEL to handle complex generative tasks, not just surface responses.

The launch of BAGEL opens up new possibilities in the field of multimodal AI, which is expected to play an important role in image editing, virtual reality, augmented reality and other fields.

OpenAI recently launched a new system, Codex, which is capable of writing, testing and debugging code in parallel using AI agents as if they were managing a virtual team of software engineers. The system is currently open for preview to Pro, Enterprise and Team users via ChatGPT.

Codex allows direct manipulation of code repositories and is sandboxed over the Internet, ensuring data security and privacy protection. This is a huge step forward for organizations that need to develop and maintain code repositories efficiently.

Key features include:

• Parallel processing: Codex can execute multiple tasks simultaneously to improve development efficiency.
• Intelligent debugging: Automatically detects and fixes errors in code, reducing the need for manual intervention.
• Easy Integration: Seamlessly interface with existing code repositories without additional configuration.

With Codex, developers can focus more on innovation and business logic design and leave the tedious code work to AI. This not only increases the speed of development, but also reduces the error rate, making software development more efficient and reliable.

Update: Mistral AI has rolled out a brand new agent feature on its chat platform Le Chat. The update not only completely upgrades the look of the agent, but also enhances its functionality, enabling the agent to use all the tools and connectors available in regular chat.

This update aims to provide a smoother and more efficient user experience. By integrating these tools and connectors, users can multitask more easily and thus be more productive.

Key improvements include:

• New interface design to enhance the visual experience;
• Expanded feature set to support more tools and connectors;
• Faster response times for a seamless interactive experience.

Mistral AI claims that this is one of the fastest agent systems on the market today. This is an important update to keep an eye on for businesses and individuals who need to communicate and collaborate efficiently.

For more information, please visit:https://t.co/FkJPXDyfxE

Google has announced the launch of MedGemma, a new product that is a collection of new variants of the Gemma 3 series designed to accelerate the development of healthcare-based AI.

MedGemma integrates the latest machine learning technologies and optimized algorithms to process medical data more efficiently. Its application scenarios are wide-ranging, including but not limited to disease diagnosis, patient monitoring and personalized treatment plan recommendation.

Key Features:

• Enhanced data processing capabilities to support large-scale medical datasets
• Improved model training speed and accuracy
• Provides a rich API interface for easy integration by developers

With MedGemma, healthcare researchers and developers can more quickly develop efficient AI solutions to improve the quality and efficiency of healthcare services.

Microsoft recently announced the open-sourcing of a complete AI adversarial security training system - theAI Red Teaming Playground Labs, designed to help users better learn and practice AI security testing.

The lab contains 12 progressive challenges covering core AI attack techniques such as hint injection, meta-hint extraction, and multi-round attacks, and builds a realistic test environment based on Microsoft's Chat Copilot. Through these challenges, users can gradually improve their AI security skills.

• Credential Theft Challenge: Learn how to obtain sensitive information through social engineering means.
• Meta-cue extraction techniques: mastering methods for mining hidden instructions from AI systems.
• Crescendo Multi-Round Attacks: practicing advanced techniques for bypassing AI security.
• Indirect Hint Injection: learn about ways to attack AI systems by modifying web page content.
• Security protection bypass: learn techniques to break through various AI content censorship mechanisms.

The lab has been designed with three difficulty levels to progressively improve users' skills from introductory to specialized. In addition, one-click deployment via Docker Compose allows you to start practicing, making it ideal for technicians and developers looking to dive deeper into AI security.

With the widespread use of AI technology, the issue of AI security has become increasingly prominent. Microsoft's open source AI Red Teaming Playground Labs provides the industry with a powerful tool that helps improve overall AI security.

You really can do it! Use any Hugging Face space as an MCP server and combine it with your local models to create endless possibilities!

In this demonstration, we generate images using FLUX and ZeroGPU techniques via Qwen 3 30B A3B, @ggml_org's llama.cpp, and @huggingface's small proxy program.

Here's a surprise.The local model is capable of accomplishing so many tasks and is able to understand and reason based on tool descriptions. This technology shows great potential, especially in automating video generation workflows and content management.

Additionally, you can access other reasoning service providers if you don't want to run it locally. Simply run npx @huggingface/tiny-agents run [TASK] Ready to go.

We also provide TypeScript and Python clients to facilitate the needs of different developers.

Sharing a simple and efficient reinforcement learning post-training framework newly released on GitHub:RL-Factorythat are specialized for the learning of intelligentsia.

The framework employs an environmentally decoupled design, where the user only needs to provide the tool configuration and reward function to start training. In addition.RL-FactoryAsynchronous tool calls are also supported, making training up to 2x faster and greatly improving development efficiency.

• Environmentally decoupled design: One-click configuration of MCP tools and custom tools;
• Asynchronous Parallel Tool Calls: Training efficiency is improved by 2x over traditional frameworks;
• Native support for the Qwen3 model: Accurate tool invocation without SFT;
• Multiple Reward Calculations: Support for rules, model judgments, and tool calls;
• Batch and distributed model judgment: Dramatically improve training efficiency;
• Multi-agent extension support: Can be converted to MCP format for interaction.

This framework is designed to make it easy and fast for everyone to train their own intelligent agents using Qwen3 and MCP tools. It is a valuable tool for developers with some basic knowledge of reinforcement learning.

Reinforcement learning is a popular field in both academic research and industrial applications. By usingRL-Factory, researchers and developers can conduct experiments and application development more efficiently, thus advancing AI technology.

🎶 Lyria RealTime is a new experimental interactive music generation model that enables anyone to create, control and perform music in real time. This innovative technology is served through the Gemini API and users can experience its demo application on Google AI Studio.

The emergence of Lyria RealTime marks another major breakthrough in artificial intelligence in music. The model combines advanced machine learning algorithms and real-time processing technology to make music creation more intuitive and flexible. Both professional music producers and amateurs can easily realize their creativity with this tool.

specifically speakingLyria RealTime supports real-time synthesis of multiple instruments, allowing users to adjust parameters such as melody, tempo, and timbre through a simple interactive interface, and instantly hear the effects of their modifications. In addition, the model also features automatic accompaniment generation, further enriching the expressive power of the music.

Currently, Lyria RealTime has been applied in several scenarios, including online music teaching, live performances, and personal creation. In the future, as the technology continues to advance and improve, it is expected to demonstrate its unique value in more areas.

Chrome v137 DevTools has a new feature that helps developers understand performance trace results with the Gemini intelligent assistant.

Specifically, Gemini automatically analyzes events in performance trace logs and generates easy-to-understand annotated descriptions. This feature will greatly enhance the efficiency of developing and optimizing performance.

It is very simple to use:

1. In the Main track of the Performance track, double-click the event you are interested in.
2. Then click the Generate label button next to the pop-up input box.
3. Gemini combines stack trace and contextual information to automatically give clear event labeling suggestions.

This feature not only simplifies the debugging process, but also enables developers to locate and resolve issues more quickly.

it's worth noting that..., While advances in AI technology have provided developers with powerful assistive tools, traditional Debug skills are still indispensable. These skills include an understanding of code logic, methods of problem troubleshooting, and an in-depth grasp of system architecture, all of which AI can't completely replace at this time.

As web applications become more complex and performance optimization becomes more important, this new feature of Chrome DevTools certainly provides developers with a powerful tool to better understand and optimize application performance during the development process.

Tesla released a clip of itsOptimusThe latest video of the humanoid robot demonstrates its ability to perform several new tasks without assistance. The robot learns by watching videos of human behavior, and all tasks are handled by a single neural network.

Tesla's humanoid robotics program aims to develop general-purpose robots capable of performing complex tasks for both domestic and industrial environments. This video release shows Optimus demonstrating a variety of abilities including lifting objects, assembling parts, and more, all skills learned by the robot on its own.

Key Technology Points::

• Single Neural Network Architecture
• Vision-based learning capabilities
• Autonomous tasking

This development marks a major breakthrough in artificial intelligence and robotics, which is expected to play an important role in various fields in the future.