The newest member of the Grok 3 family is now live on our API.
The Grok 3 Mini redefines cost-effective intelligence solutions by outperforming existing inference models by a factor of five. This small model excels in several real-world application scenarios and is especially suited for tasks that are cost-sensitive but require efficient processing.
Grok 3As the world's strongest non-reasoning model, it has demonstrated outstanding capabilities in fields such as law, finance and healthcare. With its powerful knowledge base and data processing capabilities, it is able to cope with complex real-world problems and provide users with accurate information support.
Whether it is dealing with complex clauses in legal documents or analyzing financial market dynamics, Grok 3 provides accurate and efficient solutions. In addition, its application in the medical field is also highly regarded, helping doctors to quickly access the latest medical research and improve the accuracy of diagnosis and treatment.
By bringing the Grok 3 family to APIs, we hope to provide a broad range of developers and enterprises with more powerful and flexible smart tools to help them achieve greater success in their respective fields.
---
The Meta FAIR team recently announced the release of a series of new research findings, and these recent advances are incrementally changing our understanding of perception and advancing advanced machine intelligence (AMI).
The projects in this release include:
- Meta-aware encoder: a large-scale visual coder that excels in image and video tasks.
- Meta Perceptual Language Modeling: a fully open and reproducible visual-linguistic model designed to handle visual recognition tasks.
- Meta Locate 3D: An end-to-end model for accurate object localization in a 3D environment.
- Model weights for the 8 billion parameter dynamic byte latent variable transformer were released, an alternative to traditional tokenization methods that promises to redefine standards of efficiency and reliability for language models.
- Collaborative Reasoning Framework: a framework for assessing and enhancing collaborative reasoning in language models.
These research tools not only provide valuable resources for researchers, but also bring more efficient and accurate AI systems to organizations and developers. Interested users can download the code, datasets, and related papers via this link.
---
The Google AI Developer Team announced that it is now possible to run Gemma 3 27B models on desktop GPUs. With the new QAT (Quantitative Awareness Training) optimized int4 model, VRAM requirements have been reduced dramatically, from 54GB to 14.1GB, while maintaining high quality output.
This breakthrough enables consumer graphics cards such as the NVIDIA RTX 3090 to run large-scale language models, thus providing more computing power for individual users and small businesses. Users can easily access these models through platforms such as Ollama, Hugging Face and LM Studio.
Key Features:
- Significantly reduces VRAM requirements and improves resource utilization
- Supports consumer-grade graphics cards, lowering the hardware barrier
- Multi-platform support for easy user selection
This technological advancement not only improves the accessibility of the model, but also provides more opportunities for researchers and developers to experiment and innovate.
---
OpenAI recently announced that it will support the Model Context Protocol (MCP) initiated by Anthropic. This open standard aims to connect language models with external tools and proprietary data sources to further enhance the functionality of the models and application scenarios.
The protocol will be integrated into OpenAI's Agents SDK, with plans to expand to ChatGPT desktop apps and the Responses API in the future.With MCP, developers can easily connect to more than 6,000 plug-and-play local or remote tools, drastically reducing the need for custom integrations.
This move not only improves development efficiency, but also lowers the technical barrier, enabling more companies and individuals to innovate with advanced language models.
It's worth noting that Anthropic, OpenAI, and other industry participants are working together to advance this standard, reflecting the industry's tendency to collaborate to solve common problems and avoid industry fragmentation.
---
ControlNet author Minjin has once again open-sourced a compelling new project: Framepack, an efficient video generation framework.
Using the 13B model, Framepack is able to generate 1 minute of video at 30fps with 6GB of video memory. On an RTX 4090 graphics card, it generates up to 1.5 seconds per frame. The technology generates video incrementally by using a next-frame prediction structure, allowing for up to one minute of continuous action video, even from a single image.
Key features include:
- Low video memory requirements: only 6GB of video memory is needed to run.
- High performance: generation speeds of up to 1.5 seconds per frame on RTX 4090 graphics cards.
- Wide range of applications: Suitable for a wide range of video content such as character dance, action, scene changes, etc.
Framepack is particularly suitable for fields that require dynamic visual effects, such as film and television production, animation design and virtual reality. Currently, the project provides a Windows one-click installation package with a visual operation interface, which is convenient for users to get started quickly.
---
ChatGPT has recently introduced a new memory feature called 'Moonshine' which has attracted a lot of attention.
The feature enhances the coherence and personalization of conversations with several new sections. First, a new "Model Setting Context" section has been added to the system prompt, listing stored memory entries such as the user's name and location, among other basic information.
Next is the "Assistant Response Preferences" section, which summarizes the user's specific requirements for responses based on their past usage history. For example, users prefer concise and direct responses, and do not like long explanations or apologies.
There is also a new 'Past Conversation Topic Highlights' section, which summarizes the main topics discussed by users in previous conversations. This section covers not only the user's area of specialization, such as artificial intelligence, machine learning, and AI-assisted creativity, but also the user's technical background, such as Linux system administration, web development, and cloud technologies.
The "Useful User Insights" section contains user-specific information such as the user's location and occupation. This section of information helps to further personalize the conversation and enhance the user experience.
Finally, the "Recent Conversation Content" section records a log of the user's latest conversations, including timestamps, titles, and message content. The "User Interaction Metadata" section is automatically generated from the user's device and usage pattern, providing detailed usage statistics to help optimize the conversation experience.
These new features provide users with a more personalized and coherent conversation experience through multi-dimensional information integration, significantly enhancing ChatGPT's utility and user satisfaction.
---
Anthropic's latest case study shows that code security tool Semgrep achieves significant security performance improvements by integrating the Claude 3.7 Sonnet model on the Amazon Bedrock platform.
The solution is able to confidently mark 201 TP3T of code security alerts as secure (921 TP3T and 961 TP3T of developer and researcher agreement, respectively) and improves false alarm detection by 161 TP3T and component labeling by 171 TP3T compared to GPT-4o. In addition, the technology supports advanced features such as automated vulnerability remediation and destructive change analysis features, enabling more secure and efficient applications in enterprise-class environments.
Key features include:
- Highly accurate classification of security alarms
- Reduce false alarms and improve labeling accuracy
- Automated vulnerability remediation and change analysis
This technology not only enhances code security for organizations, but also brings greater efficiency and reliability in real-world deployments.
---
OpenAI's latest o4 mini model has been significantly upgraded in terms of its code of behavior and internal work instructions, marking its evolution from a purely computational model to an intelligent agent with autonomous capabilities. This change is not only in the technology, but also in the overall improvement of its practicality and interactivity.
Compared to its predecessor model o3, o4 mini is not only capable of accomplishing tasks through the use of a variety of tools, but also capable of repeatedly verifying the accuracy of the answers during the execution process to ensure the reliability of the results and the successful completion of the task. This mechanism not only improves the usefulness of the model, but also greatly reduces the possibility of errors. In addition, o4 mini shows greater adaptability and flexibility in handling complex tasks, making it more efficient in real-world application scenarios.
For example, in areas such as customer service, data analysis and content generation, o4 mini enables more intelligent and humanized interactions to enhance the user experience. With the continuous development of AI technology, these improvements in o4 mini make it even more promising for future applications.
---
Google DeepMind has released a new Quantisation Aware Training (QAT) version of the Gemma 3 model, which significantly reduces size while maintaining performance.The Gemma 3 27B parameter version of the model has been compressed from 54GB to 14.1GB, dramatically reducing storage and compute resource requirements. This is a significant breakthrough for application scenarios that require the deployment of large-scale language models in edge devices or resource-constrained environments.
In addition, the model's open source checkpoint is now available to the public, allowing developers to test and apply it on a variety of platforms, including MLX, llama.cpp, and lmstudio. This not only improves the usability of the model, but also promotes further research and improvement in the community.
The QAT technique allows the model to better adapt to the loss of accuracy due to quantization when actually deployed by simulating a low-precision computational environment during the training process. This optimization method ensures that the model maintains high performance even after quantization, thus providing the possibility for a wide range of applications.
---
Recently, a program calledFramePack's native video generation tool has attracted a lot of attention. The tool claims to require only6GB video memoryReady to run, greatly reducing the hardware threshold.
From the example posted, theFramePackThe ability to convert input images (Figure 2) into high-quality motion video (Figure 1) is impressive. The developers have released an all-in-one installer for Windows, which makes it easy for users to use it directly; meanwhile, Linux users can also run the tool from the command line.
FramePackThe launch of the program provides a new option for content creators, video editors, and researchers. However, due to the complex image processing and video generation techniques involved, the developers recommend that users with some basic programming skills try it out, otherwise they may encounter some technical difficulties (Figure 3). For experienced users, theFramePackUndoubtedly a powerful creative tool.
