Large Model

On September 20, Chinese telecommunications released on September 17, a new-generation star-size model, Xing4.0-29B-A4B, with a total of 29B, a activation parameter of only 4B, with original support of 256K in the context, which can be extended to 512K, the first large-scale model in the country to optimize the depth of complex engineering tasks based on national capacity and framework. Officially, from training chips to reasoning deployment, the whole of the country's production — training, adaptation of the national production framework, self-study structures, open-source ecology — is one route. From Super..

On September 5th, according to AGI, the latest large model of millet has entered the internal run-off phase and will be released in September as soon as possible. This information has not yet been officially confirmed by Mi. On the same day, OpenCode Go online anonymous model Omen Alpha was opened for testing by subscribers; it was reported to be linked to MiMo V3, but there is no publicly available evidence to confirm model identity. Notifications received by users of MiMo-V2.5-Pro-UltraSpeed showed that this version..

On September 1st, the S.I.S.S.S.S.-Syllable Fire announced today the launch and opening source of X2.5-4B and X2.5-1.7B generic large models at the end. It was stated that the two models had original support of up to 1 million Token context windows and were the only products in the current end-side model that supported that level of context. The two models use a mix of attention structure and optimize around core competencies such as intelligence, code, mathematics and command compliance. Based on the National Calculator Platform, completes the whole process, using about 20 trillion Token to diversify into..

According to news from August 28th, it was reported today that, with the accelerated landing of artificial intelligence applications, our average daily word calls had exceeded 500 trillion by June 2026. Behind the rapid growth in the size of the words, the competition for the global large model has become hot, and there has been an explosion in the demand for power. According to the view of the economy, the big global model is currently at the most competitive and iterative stage, and the big China model is already in the first stage of global competition. According to industry sources, the current AI flag model is being updated almost monthly, and the focus of competition has shifted from a simple “model IQ” to an intellectual..

On 8 August, according to the Financial Times on 7 August, byte beats are training a large model of up to 10 trillion parameters and are still in the early stages of pre-training. According to three informed sources, this size will be more than three times the size of the largest model published in China, Kimi K3, the dark side of the moon, and close to the volume of Mythos 5, the most advanced model of Anthropic, estimated by industry at around 8 trillion parameters. According to the report, the project is led by by byte by the head of the Seed Foundation, reflecting the CEO's first call for original..

On the 7th of August, Ali Baba Video Generation Large Model Wan3.0 August 6, opened the public survey, generating a single video of up to 30 seconds and supporting input in document formats for the first time: doc, xls, ppt, pdf, md. A single file or link does not exceed 100 MB, 50 pages, and an upload product introduction to PPT can be used to generate a complete image with a hint. Wan3.0 also supports text, pictures, audio and video as reference, trying to keep roles, props, voices, spatial relationships and styles consistent, and can modify images, dramas and platforms..

According to the news of 5 August, the artificial intelligence industry has entered a new stage of development. In response to competition from a new generation of Chinese models (e.g. Kimi K3 and V4 Flash in DeepSeek, the dark side of the moon), leading AI firms around the world are significantly reducing the price of their products while upgrading the capacity of entry-level models. Recently this strategy was taken by OpenAI. The company reduced the price of the base front model ChatGPT 5.6 Luna by 80% (per million Token) and also reduced the middle model..

On August 3rd, OpenAI officially released the results of its phased research and development for the next generation of big internal models, Astra. The experiment showed that the model had independently extrapolated and gave complete proof of 10 long-standing mathematical difficulties without human hints, with a total of 249 pages. According to the technical report published by OpenAI, the mathematical dilemma of the Astra attack is an area that has been open for more than 10 years and with a chronic lack of progress, covering combination mathematics, algebra geometry and numerics. The model displays complex tectonic aids in the certification process..

On August 2nd, according to news news today, recently, OpenRouter, the Global Multimodel Convergence Platform, released the latest week of AI Big Model Call Checklist, the top five products being developed by Chinese companies. The top of the list was Mimo-V2.5, with a weekly call of 10.5 trillion Tokens and an increase of 12%. According to Xiaomei Mimo, the product manager of the Big Model Team, Xinyan Shu, the number of calls has risen from 1.5 trillion words to 10.5 trillion words a week, more than six times more in two months, just..

On July 25th, Anthropic officially launched yesterday a new generation of closed-source large models Claude Opus 5 and synchronised access to AWS. The new model, which is based on high value for money versus modifiable reflection slots (efort), is prominent in such tasks as programming, long file reasoning and complex business analysis. API pricing maintains token 5 dollars per million input, 25 dollars for output, at the same rate as Opus 4.8; start Fast mode response speeds up 2.5 times, double the bill. In CursorBenc..

On June 18, the Shanghai Stock Exchange issued a directive yesterday (17 June) to support the listing of high-quality, large-scale, artificial-intellectual models that have not yet generated a certain income. It is clear from the listing guidelines that the fifth set of listing standards applies to enterprises with large artificial intelligence models. The core technology products of the issuers applying for listing should fall within the scope of the support encouraged by national STI strategies and related industrial policies in the area of artificial intelligence, with the main operations being the autonomous development of large artificial intelligence models, model services or model applications. The issuer has a prominent position in the field of large models of artificial intelligence..

On 8 June, according to the Daily Economic News, according to OpenRouter's latest surveillance data, the total number of calls for the Global AI Large Model last week (1-7 June) reached 36.1 trillion Token, an increase of 13.51 TP3T over the previous week, and has risen for seven consecutive weeks. According to the data, China ' s AI megamodel was called 14.19 trillion token last week, with a significant increase in ring-to-ring ratio of 27.49% for three consecutive weeks, while the U.S. 's AI mega-model weekly call was 3.2 trillion T..

On May 22, a message from the Apple Official Public Quiz announced that the App, PC and web pages were connected to a new generation of large models Qwen 3.7-Max. Updates the thousands of questions APP to the latest version (Note: 6.9.7 and above) by clicking on the lower capsule "Qwen 3.7-Max" or by placing a drop-and-drop in the Model Selection Bar of the PC end and web-end dialogue interface, which is free of charge for the first time. According to the official presentation of the Magna Carta, Qwen 3.7-Max is committed to being the foundation of an all-powerful intelligent body -- whether it's a..

On May 22, according to the interface news sources, the Deputy Director of the Policy Research Unit of the National Commission for Development and Development (CND) and press spokesman Lee said in a press conference today that the core technical and application needs in the area of artificial intelligence are growing rapidly and that we have always insisted on system layout, division, open sharing and control. According to Lee, the CDRC promotes a broad and deep integration of artificial intelligence with all sectors of the economy and society, directs the larger model of national production to make a greater effort to adapt to the national production algorithm chip and, while maintaining rapid development, to ensure that it is self-manageable, well-developed and stable, so that all people can share the benefits of artificial intelligence development..

On May 20th, Ali thousands asked the Big Model today to officially launch Qwen 3.7-Max, a new generation of flagship models for the age of smarts, to be served through API. Qwen3.7-Max is the most comprehensive and powerful model of intelligence to date. According to the official presentation of the Magna Carta, Qwen 3.7-Max is committed to being the foundation of an all-powerful intelligent body -- It is both capable of developing and debugging codes, automating office processes, and continuing to implement autonomously over hundreds or thousands of steps of long-cycle tasks. Qwen3..

ON 19 MAY, THE SHANGHAI WEB OFFICE ISSUED AN UPDATED BULLETIN YESTERDAY ON THE STATUS OF THE SHANGHAI MARKET-GENERATED A.I.S.S.S.S.S.A.S.S.A.S.A.S.A.S.A.S.A.S.A.S.A.S.A.S.A.S.A. IT WAS DESCRIBED THAT THE WITA (SILICA MOTION) MEGA-MODEL CORE SHOULD BE APPLIED TO HUMAN ROBOTIC INTERACTIONS, ALLOWING FOR THE EMOTIONAL INTERACTION OF NATURAL DESIGNERS, AND FOR ROBOTS TO BE UPGRADED TO “SILICA-BASED PARTNERS” WITH A SENSE OF CONTINUITY AND PERSONALITY. THE WITA LARGE MODEL, WHICH IS AT THE CORE OF "INTERACTIVE INTELLIGENCE DEPLOYMENT" WILL BE APPLIED TO GUIDE PURCHASE GUIDES, SERVICE RETAIL STATIONS, ETC...

News from 13 May, in recent days, the Internet ' s easy-thinking cloud has launched the Enterprise-level Large Model Convergence Platform ThinkFlow, which integrates the 20-plus mainstream models of DeepSeek, Kimi, Qwen, MiniMax and others into the Unified Movement through standardized API interface, allowing for flexible switching between multiple models once an enterprise has completed an access. ThinkFlow provides a full link to the Token consumer visual viewer, which accurates the call cost of the large model to the single request level. In the meantime, the platform centrally hosts all API keys and supports the use of model suits..

On April 23rd, Mi announced today that the Xiaomi MiMo-V2.5 series of models had officially opened public measurements. The MiMo-V2.5-Pro and MiMo-V2.5 models are about to open the world. 1AI was informed from the bulletin that the Xiaomi MiMo-V2.5 series consists of MiMo-V2.5, V2.5-Pro, V2.5-TTS Series, V2.5-ASR, with a stronger reasoning, a more stable Agent, a longer context, stronger command compliance and vague command understanding, and..

On March 31st, the Agent Evaluator, ClawBench, released yesterday ' s updated list of large models covering 30 complex Agent missions covering five core business scenarios of office collaboration, information retrieval, content creation, data processing and software engineering. This list includes over 40 major mainstream models, and the top 10 of the world's top four national production models, i.e., spectra, byte, and millimetres. GLM-5-Turbo tops 93.9 points of the CLAW SCORE, making it the most comprehensive..

On 2 March, the Treasure Snippets released an open-source model for the vertical field of wind management, Ostrakon-VL, dedicated to catering services and retail outlets, and announced that the relevant technical capabilities were open to industry-wide free of charge. It has been learned that, with the help of a large “Bashizawa” model, industry participants, such as off-sale platforms, restaurants and retail enterprises, have been able to quickly improve their ability to identify and govern non-compliant scenes through artificial intelligence technology, in terms of image recognition, post-cook warning, etc., and significantly improve the efficiency of digital governance. According to Chinese Internet technology, in the area of smart inspection of Ming Zai, "Bashizawa" can take 24 hours..

On February 11, an ants group open source released a large full-mode model Ming-Flash-Omni 2.0. In a number of open benchmarking tests, the model has been prominent in key competencies such as visual language understanding, voice-controllable generation, image generation and editing. Ming-Flash-Omni 2.0 was described as the first industry-wide unified audio generation model to generate both voice, environmental sound and music in the same track. Users can exercise precision control over sound, speed, tone, volume, emotions and dialects by using only natural language instructions. Model..

According to news from the South China Morning Post, on 3 February, during the last sprint before spring, several forward artificial intelligence laboratories in the country are intensively launching a new generation of large models to try to capture exposure and user intelligence during the festival. Ali and the Dark Side of Moon were also exposed to updating the flagship model in the next two weeks after the first release of Qwen3-Max-Thinking and Kimi 2.5 last week. According to the source, AI plans to launch the GLM-5 before the Spring Festival, which is the fifth-generation version of the GLM series, which is expected to..

In January 10th, The Information reported that DeepSeek would launch a new flagship AI model - DeepSeek V4 - in mid-February this year during the new calendar year, and would be better able to write codes. Internal tests indicate that its AI programming performance is expected to outperform industry leaders, including OpenAI GPT and Anthropic Claude. According to the source, DeepSeek V4 is here..

On November 14, New Wave Weibo released its first large open source model, VibeThinker-1.5B, called “Small models can also be intelligent”. 1AI is accompanied by the following official presentation: Are the most powerful models in the industry at present more than 1T, or even 2T-scale models, that have a high level of intelligence? Is only a small number of technology giants capable of making large models? VibeThinker-1.5B, which is exactly the negative answer given by Weibo AI to this question, proves that small models can also have high IQs. It's..