On November 13, 2025, OpenAI officially went live with the GPT-5.1 API, with comprehensive performance improvements, including faster response times, enhanced control flexibility and optimized code generation capabilities. Baidu released ERNIE 5 model, which surpasses GPT-5 in chart comprehension, document parsing and multimodal processing, etc. BentoML supports trillion-parameter open-source models Ling-1T and Ring-1T, PyTorch and LAMMPS integration accelerates material science simulation, and IBM Granite Nano model enables cell phone offline AI operation.
- This summary is generated by AI analyzing the content of the article for informational purposes only.



Major updates include:
- Adaptive reasoning, adjusting depth of thought based on task complexity
- "No inference mode" for instant replies
- Reduced number of tokens required for simplified tasks
- Deeper thinking under complex mandates
- 24-hour prompt caching for improved responsiveness
- Further improvement in code quality
- Improved user-directed update messages (preambles) in tool call sequences
OpenAI announced GPT-5.1 is now officially available in the API.
This new version of the model not only offers significant speed improvements, but also enhanced controllability and better performance in code generation. In addition. GPT-5.1 also comes with a range of useful new tools designed to provide more comprehensive support for developers.
Key improvements include:
- Faster response times and reduced wait times
- More flexible control options for easy customization of outputs
- Stronger code generation capabilities to improve programming efficiency
- New utilities to expand application scenarios
Whether it's developing complex software systems or creating efficient automation solutions, the The GPT-5.1 can all bring significant improvements to your program.
Recently, Baidu Inc. announced the launch of its latest natural language processing modelERNIE 5, the model outperforms the widely publicized GPT-5 in several key performance metrics.
According to the test results, ERNIE 5 performs well in chart comprehension, document parsing, and multimodal information processing, especially when dealing with complex documents and cross-modal data, showing significant advantages.
Baidu says ERNIE 5 uses advanced pre-training techniques and large-scale datasets to better understand and generate high-quality text content. In addition, the model has strong cross-linguistic capabilities, supporting processing tasks in multiple languages.
- **Chart Comprehension**: ERNIE 5 is able to accurately extract key information from charts and graphs and analyze them effectively.
- **Document comprehension**: ERNIE 5 performs particularly well when dealing with long documents and complex formatting.
- **Multimodal Information Processing**: Combining image, audio, and other multimodal data, ERNIE 5 is able to provide a more comprehensive understanding of information.
With the release of ERNIE 5, Baidu further consolidates its leading position in the field of artificial intelligence and provides stronger technical support for various industries.
GPT-5.1 is now fully available in the API, with pricing aligned with GPT-5.
The update also introduces The gpt-5.1-codex and gpt-5.1-codex-mini versions are designed to handle long-running programming tasks. These new models improve developer productivity by enabling more efficient generation of complex code through deep learning techniques.
Major updates:
- gpt-5.1-codex: for large code projects, providing more powerful code generation capabilities.
- gpt-5.1-codex-mini: lightweight version for smaller scale code writing.
Additionally, we have improved the hint caching mechanism, and the validity of the cache is now extended to 24 hours to further enhance the user experience. Detailed technical reviews and usage guidelines have been published in our blog posts.
Magentic Marketplace is an open-source simulation environment designed to study the interactions and transactions of AI agents in digital markets. Its modular design allows researchers to experiment with different market settings, agent behaviors, and system dynamics at scale.
The platform provides a wealth of tools and resources to help researchers build and test various market models. By using magnetic markets, researchers can gain insight into the behavioral patterns of AI agents under different conditions to optimize market mechanisms and improve trading efficiency.
Magnetic Markets not only supports basic market simulations, but also allows users to customize complex economic models, including auctions, bidding, and other trading mechanisms. In addition, it integrates advanced data analysis and visualization tools to help researchers better understand and interpret experimental results.
application scenario::
- Research on market behavior and dynamics
- Testing new trading mechanisms and algorithms
- Optimizing e-commerce and online marketplace operations
- Education and training for teaching market economics and artificial intelligence
Microsoft Research has deployed the Magnetic Marketplace on Azure AI Foundry Labs and welcomes researchers and developers to explore and use it further.
For more details, please visit: https://t.co/gwCfDS8rkP
LlamaParse recently launched a major upgrade, a new experimental "Proxy Chart Parsing" feature designed to significantly improve the accuracy and efficiency of chart OCR. This innovative technology solves the problem most large language models face when converting charts to accurate numeric data by tracing the contours in a line chart and extracting the values.
While traditional chart OCR methods often require a great deal of manual correction work, LlamaParse's new features can drastically reduce this burden and make it possible to automate chart analysis. Whether you are a financial analyst, a researcher or a business decision maker, you can use this tool to quickly extract and analyze key information from charts and improve your productivity.
Users can now experience this new feature on the LlamaParse platform. Simply upload a chart and the system will automatically parse it and provide accurate data output.
BentoML now fully supports the latest trillion-parameter open source models Ling-1T and Ring-1T from AntLingAGI, marking the official entry of the era of large-scale language modeling into the open source and production phases.
**The time has come for trillion-parameter models** that are not only open source but also efficient for real-world production environments. For users exploring open source models, Ling-1T and Ring-1T are excellent choices.
- **Ling-1T (non-thinking)**:1 trillion total parameters, about 50 billion active parameters (MoE, 1/32 active)The newest version of YaRN supports up to 128K contexts (YaRN), with efficient reasoning capabilities: high accuracy and fewer tokens required, making it especially suited for code, math, and logical reasoning tasks.
- **Ring-1T (Thinking) **:1 trillion total parameters, about 50 billion active parameters (RLVR)The company has been stabilized by Icepop and Trillion Scale ASystem with leading performance or competitive in benchmarks such as AIME/HMMT, LiveCodeBench, ARC-AGI, etc., with silver-level International Mathematical Olympiad reasoning, and similarly supports up to 128K contexts (YaRN).
BentoML provides out-of-the-box support for deploying and scaling these models on your GPU cluster or any cloud platform. For example, an FP8 variant can be run on 8 B200 GPUs on a single node to reduce cross-node communication overhead.
If you need a customized deployment solution, please contact us.
By integrating PyTorch-based machine learning force fields (MLIPs) into the LAMMPS Molecular Dynamics (MD) package, scientists are able to achieve fast and scalable simulations of atomic systems, which are important in chemistry and materials science research.
Using the ML-IAP-Kokkos interface, developers can simplify the process of interfacing community models and massively scale MD simulations. This technology not only improves computational efficiency, but also opens up new possibilities for modeling complex material systems.
Key benefits include:
- **Increased computation speed**: Significant reduction in simulation time through GPU acceleration.
- **Enhanced scalability**: Supports massively parallel computing for larger, more complex systems.
- **Simplify the development process**: provide a unified interface to lower the development threshold.
🔗 https://t.co/Bw7d0QrdFO
#PyTorch #OpenSourceAI #AI #Inference #Innovation
In addition to PyTorch itself, the PyTorch Foundation hosts several important projects such as vLLM, DeepSpeed, and Ray, demonstrating that PyTorch is becoming a foundational language in the field of artificial intelligence.
For Luca Antiga, CTO of Lightning AI, who also chairs the inaugural PyTorch Foundation Technical Advisory Committee (TAC), the PyTorch Conference is an important opportunity to connect with the community that is driving PyTorch forward. This community includes millions of developers who develop with PyTorch Lightning.
The event is notable because you have the opportunity to meet developers who have built amazing applications based on PyTorch. These applications cover a wide range of fields from natural language processing to computer vision, demonstrating the power of PyTorch in real-world applications.
The conference not only facilitated technical exchanges, but also provided valuable opportunities for the open source community to interact. Through events like this, PyTorch further solidifies its central position in the AI ecosystem.
IBM's Granite Nano model now lets users run AI completely offline on their phones.The technology makes complex AI tasks achievable on mobile devices by optimizing model size and computational efficiency.
Emma, a member of IBM's Granite team, introduced this innovative technology in detail. By using the Granite Nano model, users can perform tasks such as image recognition and voice processing on their cell phones without relying on an Internet connection, which greatly enhances the intelligent application capabilities of mobile devices.
Key features include:
- **LOW POWER CONSUMPTION**: The Granite Nano model is designed with the battery life of mobile devices in mind, ensuring long hours of operation.
- **HIGH PERFORMANCE**: despite the small size of the model, performance is not compromised and complex tasks can be handled quickly.
- **Easy to Integrate**: Developers can easily integrate these models into existing applications to enhance the user experience.
The potential value of this technology can be seen in a wide range of applications, from smart homes to healthcare.


Comments are closed.