Google Research recently announced the innovative 'Titans' series model architecture, achieving a groundbreaking 2 million token context length through bionic design, with plans to open-source related technologies in the future. The core innovation of this architecture is the introduction of a Deep Neural Long-Term Memory Module, inspired by the human memory system. Titans cleverly combines the rapid response capability of short-term memory with the persistence characteristics of long-term memory, while utilizing an attention mechanism to handle immediate context, forming an efficient information processing system.

In the field of artificial intelligence, ByteDance's commercial technology team presents the latest achievement, the Infinity model, which has become a new leader in the auto-regressive text-to-image generation domain with its outstanding performance and innovative technology. This newly open-sourced model not only surpasses Stable Diffusion 3 in image generation quality but also demonstrates significant advantages in inference speed. The core innovation of the Infinity model lies in its use of a Bitwise Token auto-regressive framework, which predicts the next level of resolution by +1 or -1.

On December 18th, at the Volcano Engine FORCE Power Conference, Volcano Engine President Tan Dai unveiled the new Doubao Visual Understanding Model. This model processes both text and image-related information simultaneously, providing more accurate answers. The new model excels in content recognition, understanding, and reasoning capabilities, featuring more nuanced visual description abilities, which brings stronger intelligent processing capabilities to enterprise users. Tan Dai revealed at the conference that the Doubao Visual Understanding Model has achieved significant technological breakthroughs and has also seen a dramatic reduction in pricing.

In the current field of artificial intelligence, precise control over language models has become crucial for developers and data scientists. Anthropic's Claude language models provide users with numerous possibilities, but effectively managing token usage remains a challenge. To address this issue, Anthropic has launched a brand new token counting API aimed at providing deeper insights into token usage, thereby enhancing the efficiency and control of interactions with language models. Tokens play a role in language models.

Elon Musk's xAI has finally fulfilled its promise from August by officially launching the API interface for its flagship AI model, Grok. However, this long-awaited API version feels somewhat rushed and basic. Currently, the xAI API only offers a model named "grok-beta." In terms of pricing, input tokens are charged at $5 per million tokens (approximately 750,000 words), while output tokens are charged at $15 per million tokens.

Zhipu AI announced today that since being showcased at KDD 2024, its flagship language model GLM-4-Plus and the first dynamic video understanding model GLM-4V-Plus have received widespread recognition, achieving outstanding results on authoritative evaluation lists, breaking the monopoly of foreign models in the top three and positioning themselves at the international elite level.

iFlytek has announced a major upgrade to the Spark API platform. This upgrade includes the Spark Max model and the 4.0 Ultra model, which have improved generation speeds by 70%, while also achieving significant breakthroughs in key performance areas such as logical reasoning, creativity, and online search. The upgraded Spark Max and 4.0 Ultra models demonstrate faster speeds and greater accuracy in logical reasoning, especially in handling complex tasks, providing detailed reasoning processes.

Kimi Smart Assistant, a subsidiary of Beijing Yue's Dark Side Technology Co., Ltd., recently announced that its Kimi API now supports online search functionality. This marks a significant enhancement in the interactive experience, allowing users to retrieve information from the Internet via the API. This update to the Kimi API enables developers to leverage this feature, allowing applications to 'surf the web' by automatically performing online searches and providing answers when encountering questions during conversations.

NVIDIA has collaborated with several universities to introduce NVEagle, a large visual language model capable of chatting using images. NVEagle can analyze image content and provide accurate answers, such as identifying individuals in images, like Jensen Huang. The model significantly enhances the understanding of visual information by transforming images into visual tokens and combining them with text embeddings. In addressing the challenges of high-resolution image processing, the research team has constructed models like Eagle-X5-7B and Eagle-X by exploring various visual encoders and fusion strategies.

Magic, a startup focused on generative AI programming, has announced the completion of $320 million in funding, bringing its total financing to nearly $465 million. This round of funding was led by former Google CEO Eric Schmidt, with participation from several well-known investors. Magic's core technology, the Long-Term Memory Network (LTM), can handle vast amounts of data and avoid bias when generating code. Its latest model, LTM-2-mini, can process 100 million tokens, equivalent to 10 million lines of code or 750 novels. In partnership with Google Cloud, it plans to establish two '

Google has launched the Gemini API's Gemini1.5Flash free plan, offering 1.5 billion free tokens every day, including 15 requests per minute, 1 million tokens per minute, and 1 million tokens per hour of context caching services, along with fully free tuning capabilities. This initiative aims to lower the barriers for developers to utilize AI technology, supporting the development of innovative applications. For small and medium-sized development teams, this will significantly reduce costs and provide convenience in R&D. This move is expected to attract more developers, driving growth.

OpenAI has announced that third-party developers can fine-tune the newly introduced multimodal large model GPT-4o. This feature allows developers to adjust the model's behavior based on specific needs, such as changing the tone, following instructions, or improving the accuracy of technical tasks. Fine-tuning can achieve significant performance improvements with only a small amount of training data and is easy to operate. OpenAI is providing developers with a free allowance of 1 million tokens daily until September 23, as well as a daily free allowance of 2 million tokens for the GPT-4o mini model.

Zhipu AI announces that the LLM GLM-4-Long, which supports ultra-long context lengths, is now available on the open platform bigmodel.cn. This model is designed specifically for processing long texts and can read the equivalent of two copies of 'Dream of the Red Chamber' or 125 papers in one go. It is widely applied in scenarios such as translating long documents, global analysis of financial reports, extracting key information, and building chatbots with ultra-long memory.