We announce the release of the new 8 billion parameterDynamic byte potential converter(Dynamic Byte Latent Transformer) for modeling weights. The model is a new technique that replaces traditional word-splitting methods and has the potential to redefine the efficiency and reliability of language modeling.
While traditional word-splitting methods have many limitations when dealing with natural languages, the Dynamic Byte Potential Transformer improves the flexibility and performance of the model by directly processing the byte stream. This approach not only better handles rare vocabularies and multilingual texts, but also significantly reduces storage and computational resource requirements.
The released model weights and related code are open for download, allowing researchers to further explore their performance in a variety of application scenarios, including but not limited to machine translation, sentiment analysis, and dialog systems.
For more information on the Dynamic Byte Potential Converter and the breakthroughs it could lead to, see our research paper and visit the links below to download the model and code:https://t.co/GgJEPXTH8WThe
---
Artificial intelligence is revolutionizing the advertising industry, and the results are astounding.
As technology continues to advance,** AI virtual spokespersons** have been able to hold any product and say anything you want. At the heart of this technology lies deep learning and natural language processing, allowing avatars to be not only lifelike in appearance, but also voice and expression like real people.
This change has given advertisers unprecedented flexibility and creative space. Brands can quickly respond to market changes through customized virtual spokespersons while saving a lot of time and cost. In addition, the emergence of virtual spokespersons also solves various problems that may be brought about by real-life endorsement, such as image crisis and contract disputes.
application scenarioIncluding but not limited to:
- Social Media Promotion
- Live online banding
- TV commercial
- Exhibition
These applications not only enhance the interactivity and attractiveness of advertisements, but also provide businesses with a wider range of channels to reach consumers.
While the technology is promising, it has also sparked some ethical and privacy discussions. How to protect the rights of consumers while utilizing AI to bring business value will be an issue to focus on in the future.
---
Recently, Tencent's newest model, Hunyuan-Turbos, performed well across multiple reviews, jumping to eighth place overall (13th for style control), right behind Deepseek-R1. The model was ranked in the top ten in major categories such as Hard, Coding, and Math.
Hunyuan-Turbos has significantly improved from its February version (ranked 21st at that time), thanks to its unique technical architecture. As the first hyperscale hybrid Transformer-Mamba MoE model, Hunyuan-Turbos combines Mamba's efficient long sequence processing capabilities with Transformer's powerful contextual understanding. This combination makes it outperform GPT-4o-0806, DeepSeek-V3, and open-source models on tasks such as math, reasoning, and alignment. It also excels on knowledge-based tasks such as MMLU-Pro.
Key Features:
- Compared to the previous Turbo model, the inference cost of Hunyuan-Turbos is reduced by about 1/7.
- Math, programming, and reasoning skills were enhanced by the introduction of a slow-thinking mechanism.
- Enhanced alignment and proxy execution after fine-grained instruction tuning.
- Optimized for English language training to improve general performance.
Upgraded reward system:
- Rule-based scoring and consistency validation.
- Code sandbox feedback to improve accuracy in STEM fields.
- Generative reward mechanisms that improve quizzing and creativity and reduce reward manipulation issues.
This series of technological innovations and optimizations not only demonstrates Tencent's strong strength in the field of AI, but also provides a new direction for the future development of AI.
---
OpenAI recently launched an open-source benchmarking tool called HealthBench, designed to evaluate the performance and safety of large language models in healthcare. The benchmark consists of 5,000 multi-round conversations comprising standardized scoring guidelines written by physicians and contains 48,562 unique assessment points.
This benchmark demonstrates the steady progress of large language models in healthcare, from 161 TP3T accuracy in GPT-3.5 Turbo to 321 TP3T in GPT-4o, with even more rapid improvements recently, with GPT-4o3 scoring 601 TP3T.Additionally, smaller models such as the GPT-4.1 nano, which costs only a quarter of the cost of GPT-4o, yet it achieves a superior performance.
HealthBench The launch not only provides a standardized evaluation platform for researchers, but also provides strong support for promoting the application of large language models in the medical field. Through rigorous evaluation criteria, HealthBench can help developers better understand the strengths and weaknesses of their models to further optimize their performance and safety.
---
Congratulations to our friend @PrimeIntellect on the release of INTELLECT-2, a 32 billion parameter model trained by globally distributed reinforcement learning.
The focus of this release is not on the models themselves, but on the infrastructure they have built, a facility that will enable distributed intelligence integration to advance AI technology. The infrastructure will not only be able to support the training of large-scale models, but will also be able to efficiently distribute computational resources across geographic locations.
Key highlights include:
- Detailed technical reports provide in-depth analysis of the model architecture, training process, and performance evaluation.
- Open source code that allows researchers and developers to download and use the INTELLECT-2 model checkpoints.
This globally distributed approach to reinforcement learning has the potential for a wide range of applications in several fields, such as autonomous driving, smart healthcare, and robotics, opening up new avenues for future AI applications.
---
Recently.OpenAIAnnouncing the addition of the ChatGPTSharePointrespond in singingMicrosoft OneDriveas new data connectors. These new features will make it easier for users to access and utilize the files and data stored on these platforms for more in-depth research.
These connectors are currently available to Plus, Pro and Team subscribers, excluding users in the European Economic Area (EEA), Switzerland and the UK. This initiative is designed to further enhance ChatGPT's functionality so that it can better support the data processing needs of both business and individual users.
By integrating SharePoint and OneDrive, users can easily use documents, forms, and other files with ChatGPT to improve productivity and research quality. This is especially important for business users who need to analyze and organize large amounts of data and information.
Additionally, this update marks OpenAI's continued expansion of its ecosystem to meet the needs of users worldwide. In the future, more similar integrations and features are expected to be rolled out to further enhance the utility and flexibility of ChatGPT.
---
from multimodal processing to inference to mixed expert models (MoEs) and benchmarking.This blog.provides you with a comprehensive guide to Visual Language Models (VLMs).
Visual Language Models (VLMs) have demonstrated strong capabilities in image understanding and text generation, and are widely used in tasks such as image captioning, visual quizzing, and cross-modal retrieval. Through multimodal learning, VLMs are able to better understand the relationship between images and text, thus improving overall performance.
- Multimodal fusion: combining image and text information to improve the comprehensive understanding of the model.
- Reasoning ability: enables the model to not only recognize objects in an image, but also to understand their context and logical relationships.
- Mixed Expert Models (MoEs): Enhance the expressive and generalization capabilities of the model through the collaborative work of multiple expert submodels.
- Benchmarking: Provides a series of standardized datasets and evaluation metrics to help researchers assess and compare the performance of different models.
---
Household chores have always been a headache for many people, but with the advancement of technology, household robots are expected to solve this challenge.
In recent years, robotics has made significant progress. From sweeping robots to intelligent cooking assistants, these devices have gradually become part of family life. It is expected that in the next few years, more advanced household robots will enter the market, which can not only fulfill daily cleaning tasks, but also perform more complex operations, such as item handling and elderly care.
technological innovationis a key driver of this change. Advances in artificial intelligence, machine learning and sensor technology have enabled robots to better understand and adapt to the home environment. In addition, with the spread of IoT technology, home robots will seamlessly integrate with smart home systems to provide smarter services.
Despite its promising prospects, the widespread use of household robots still faces a number of challenges, including issues of cost, safety and privacy protection. However, these issues are expected to be gradually resolved as technology continues to mature and market demand grows.
The arrival of household robots will not only greatly reduce the burden of household chores, but will also bring more convenience and comfort to the family.
---
The temporal and synchronous computational properties of brain neurons are often overlooked in modern neural networks. However, these properties are crucial for the flexibility and adaptability of biological intelligence.
To address this problem, researchers have proposed a new neural architecture, Continuous Thought Machines (CTMs).CTMs are designed from the ground up and utilize neural dynamics as a core representation, which allows for natural adaptive computation.
Key features of CTMs include:
- The maze problem is solved by observing raw maze images and generating step-by-step instructions directly from the neural dynamics.
- In image recognition tasks, CTM naturally examines different parts of an image in steps, thus improving accuracy and interpretability.
- The CTM saves energy by adjusting its thinking time according to the difficulty of the task. For example, when recognizing a gorilla, the CTM's attention moves from the eyes to the nose to the mouth, much like human visual attention patterns.
This study emphasizes the synergy between neuroscience and artificial intelligence. Although modern AI is based on brain-inspired ideas, the two fields often operate independently. By drawing on biological inspiration and iteratively tracking emergent behaviors, researchers develop models with unexpected capabilities, such as excellent calibration performance in classification tasks.
In the future, researchers will continue to explore this novel architecture and further introduce more biological concepts to discover more exciting emerging behaviors that push the boundaries of AI technology.
---
Recently, a study calledCollaborative Reasoning (Collaborative Reasoner)framework is proposed to improve the collaborative reasoning capabilities of language models in multi-intelligent body environments. The framework lays the foundation for the development of social agents capable of cooperating with humans and other intelligences.
With the development of AI technology, more and more application scenarios require efficient interaction and collaboration among multiple intelligences, such as autonomous driving, smart home systems, and complex human-computer interaction tasks. By introducing more advanced reasoning mechanisms, the collaborative reasoning framework enables language models to not only understand individual commands, but also work with other intelligences to accomplish tasks in dynamically changing environments.
specifically speaking, this framework provides a new training methodology that enables models to learn how to effectively communicate and make decisions with different types of participants in simulated real-world scenarios. In addition, it supports many types of task settings, including but not limited to areas such as dialog management and strategy planning.
The researchers say the work will help drive the construction of smarter, more adaptable social AI systems that can better serve the needs of modern society.
