AI Models

According to news from January 21, early in the month, DeepSeek will launch a new-generation flagship AI model — DeepSeek V4 — in mid-February this year. On 20 January, on the first anniversary of DeepSeek-R1, developers discovered that DeepSeek had updated a series of FlashMLA codes in GitHub, across 114 files 28 references to unknown '...'

On 12 January, according to the China Weather Service, China's first financial weather AI model, "stock", was launched in Guangdong on 11 January. The Molybdenum Machine was developed by the University of Jordan in conjunction with China's National Meteorological Information Centre, with the aim of exploring the role of meteorological factors in the pricing of financial assets and providing innovative tools for risk management and investment decision-making. Zhao Yanxia, Director, Financial Weather Focus Open Laboratories, China Meteorological Administration, presented the financial weather model Al Al, which reanalyzes data and..

On January 6th, according to The Inside of Business, the data label company Invisible Technologies CEO Matt Fitzpatrick stated in the podcast 20VC that a core misunderstanding surrounding the generation of AI training was gradually being corrected: the role of humans in data production and feedback would not be replaced by synthetic data in the coming decades. Fitzpatrick points out that there is a long-standing view within the industry that AI..

On January 6, in the CES 2026 keynote address held today, Chief Executive Officer Hoang In-hoon of Inweida delivered a keynote speech announcing a large-scale expansion of his open-source model bank, the release of new models and data sets covering the four main areas of language, robotics, autopilot and medicine, and further acceleration of industry-wide AI innovation. Weeda contributed to the Open Source Training Framework and the world's largest open multi-modular data set, including 10 trillion language training tokens, 500,000 robot tracks, 455,000 protein structures and 100 TB vehicle sensors..

On 25 December, in a message from the Permanent Representative, the Qwen3-TTS family announced the introduction of two new AI models, Qwen3-TTS-VD-Flash and Qwen3-TTS-VC-Flash. 1AI with the following main features: Sound creation: Qwen3-TTS-VD-Flash supports the input of complex natural language commands, achieves fine-tuning of sound, rhythm, emotion, man-made, etc., and achieves full control from what to what to say, allowing users to define freely what they want..

4090,6GB. Visible notebooks can generate commercial posters with the text! Z-Image, as an efficient, light-quantitative, generating AI model, not only argues at a fast pace, but also supports bilingual understanding and precision in Chinese and English. This paper will download from the model the configuration ComfyUI to write the hints to solve common errors, hand-held to complete Z-Image's local deployment and combat, little white friendly. According to your equipment, select the deployment plan before you do it, please confirm your computer configuration. You can check the watch, quickly..

On December 3rd, The Information reported that OpenAI was developing a new AI Large Language Model to respond to the technical competition of Google Gemini 3, codenamed Garlic. It was described that the new AI model, presented to the team last week by Mark Chen, Chief Research Officer of OpenAI, was said to have performed better than Google Gemini 3.0 and Anthropic Opus 4.5 in the programming and logical reasoning mission. Open..

On November 28th, according to the British Financial Times, China has surpassed the United States in the global market for open-source artificial intelligence, i.e., the AI Model, and has gained a key advantage in the global application of this powerful technology. According to reports, a study carried out by the Massachusetts Institute of Technology and Open Source Artificial Intelligence Starter, Hugging Face found that, over the past year, the share of the open source AI model developed by the Chinese scientific team in the global download of open source models had risen to 171 TP3T, more than 15...

On 26 November, the Financial Times reported that China had surpassed the United States in the global “open source” AI market, thus gaining a key advantage in the global application of this powerful technology. A study by the United States MIT and Open Source AI Entrepreneurship, Hugging Face, found that over the past year, China ' s share of newly developed open source models in total downloads of open source models had risen to 17%, surpassing United States companies like Google, Meta and OpenAI by 15.8%. This is the first time that Chinese companies have surpassed their American counterparts in this area. Open..

On November 22, according to news, several social media posts and articles recently claimed that Google's changed policy was to use user Gmail messages and attachments to train the AI model to "blow the green light" and that the only exit method was to disable the "smart function" of a type of spelling. According to the media, The Verge today, Google began to clarify misunderstandings about its policy changes. Google spokesman Jenny Thomson said, "These stories are not accurate -- we haven't changed any settings, Gmail intelligence has existed for many years, and we won't use the user's G..

On November 19th, in the early morning hours of Beijing, Google officially launched a new Gemini 3 series of models and positioned them as the “most intelligent” “facts of fact” AI system to date. This version is also seen as a key opportunity for Google to regain its lead position in the consumer-oriented AI competition, along with the down-line GPT-5. All Google users can directly use the new flagship model Gemini 3 Pro in the Gemini application and introduce it in the search subscription service simultaneously. Deep..

On November 16, according to the news, Golem yesterday, November 15, published an article in which it reported that Google, through its Al Studio platform, was testing an unnamed AI model that was close to human experts in deciphering hard-to-read historical manuscripts. 1AI cites the presentation that the performance of the model was systematically tested by historian Mark Humphreys using a specially developed baseline data set. The results show that the overall character error rate of the model was about 1.7..

On November 7, according to news, the latest battles in the US stock version of the AI trade competition Rockalpha show that the national production model is strong, that the soybean bag is at a rate of 7.09%, and that Mini Max M2 and Kimi K2 are second and third, and that the national production model is the top three. The RockAlpha Trading Competition, launched by the Singapore Finance and Technology Corporation RockFlow, brings together mainstream national and international actors such as GPT-5, Grok, DeepSeek, Claude, Gemini, Qwen, AI..

On November 7th, the dark side of the moon launched Kimi's most powerful open-source thinking model - Kimi K2 Thinking. The model was described as a new generation of Tinking Agent, trained in the concept of the "model is Agent" on the dark side of the moon, where originals mastered the ability to "think and use tools". In the final human examination (Humanity & #039; s Last Exam), autonomous web browse Comp, complex information gathering reasoning (SEAL-0)

On 6 November, according to the news, ufukuzen, a tech-media ufukuzen, released an article yesterday (5 November) reporting that on the Google Enterprise Platform Vertex AI, he found traces of his next-generation flagship AI model Gemini 3 Pro preview, labelled "gemini-3-pro-pre-review-11-2025 " , which is expected to be released in November 2025. From model labels and available exposure information, Google predicts that it will be open to selected corporate clients in November 2025..

On 4 November, according to Google, on 1 November, after a Republican senator complained about her development-oriented AI model Gemma “fabricated serious criminal charges against her”, the company had taken the model from the AI Studio platform. 1AI notes that the official Google News account states in a post on platform X that the company “takes note of reports that non-developer users have attempted to use the Gemma model in AI Studio and raises factual questions with it”. AI Studio is a developer--

On 3 November, in the news, Ali Thong Yi asked about the latest flagship model, Qwen3-Max, which is already “deep thinking” on the official web line, a feature that significantly enhances the efficiency of complex task management by enhancing the ability of the chain of reasoning to analyse and dismantle multi-step issues. 1AI noted that Aliyun launched Qwen3-Max on September 24, which is the largest and most capable language model of a communism team to date, with parameters exceeding 1 trillion dollars and pre-training data as high as 36T tokens. Qwen3-Max-Instract Official Edition on behalf of..

News from 3 November, on 1 September, the United States officially released the LongCat-Flash series model, which is now available in two versions, LongCat-Flash-Chat and LongCat-Flash-Thinking, and has received the attention of developers. Today, the LongCat-Flash series officially publishes a new family member, LongCat-Flash-Omni. 1AI was informed by the official introduction that LongCat-Flash-Omni was using LongCat..

On October 28th, a day ago, Google announced the expansion of its "Earth AI" capability on Google Earth and Google Cloud platforms to launch a new generation of geospatial AI models and reasoning intelligence. The system combines underlying models with multi-state reasoning and is designed to provide viable recommendations for complex global issues. Google states that the core of Earth AI is to combine multi-area models such as images, population and environment with reasoning Agent. Agent can break down the problem of natural languages into a multi-step plan..

On October 17th, the Google Institute (15 October) published a book announcing that it was working with Yale University to jointly publish a Cell2Sentence-Scale 27B model based on Gemma to help identify new potential cancer treatments by analysing single cells. 1AI cites Boven as one of the major challenges in the case of cancer immunisation therapy is the “invisibility” of many tumors to the immune system, known as “cold” tumours. A key strategy is through an antigen prese..

On October 16, the Anthropic official network today (16 October) published a blog announcing the launch of the small AI model Claude Haiku 4.5. Officially, the model provides a powerful performance near the front line at very low cost, and the target is an ideal option for real-time, low-delayed tasks such as chat assistants and customer service. Briefly describe the three sizes of the Anthropic Claude series model: Haiku (small), Sonnet (medium) and Opus (large). Big model..

On October 15th, BGR published yesterday, October 14th, an article in which it reported that Google is expected to release the next generation of AI Model Gemini 3.0 on October 22, which will mainly improve delays, costs and coding and significantly enhance the performance of SVG image generation. After OpenAI released GPT-5 in early August, Anthropic launched Claude Sonnet 4.5 at the end of September, competition in artificial intelligence became increasingly hot, and industry focused on another..

On October 14, according to Futurism, the University of Stanford's latest paper indicates that when artificial intelligence models are rewarded with interactive indicators such as "shows" in social media settings, there is a growing tendency towards "antisocialization". The research team tested a variety of AI models in three scenarios, namely, the simulation of elections, the sale of commodities and social media, including Qwen in Aliyun and Llama in Meta, and found that, even when protective measures were put in place, the models could be distorted and misbehavioured by competition. Co-author of the thesis, Professor of Stanford University, James Z..

According to October 13, according to a foreign media source, TechXplore, recent studies by Anthropic, AI UK and Alan Turing found that even the largest AI model, with about 250 malicious files, could be successfully invaded. Training data on large-language models are mostly derived from open networks, which allow them to accumulate a vast knowledge base and generate natural languages, but are also exposed to the risk of data poisoning. It was generally accepted in the past that, as the size of the model grew, the risk would be diluted because the proportion of poisoning data would need to remain constant..