Model

News from September 9, Mitech announced on September 8 that the Xiaomi Mimo Desktop desktop client was open for invitation. Users who apply for pass-through can use MiMo-X-Pro-Pro-Preview and MiMo-X-Flash-Preview for free during the test period. Mimo Desktop receives materials in various formats, breaks targets and calls tools, and finally delivers results such as PPT, Office files, web pages and lightweight games that can continue to be edited. It supports..

On September 2nd, the "AI godmother" Lee Feife's founder, World Labs, released today the "first multi-modular world model of the world" -- Atlas, claiming that the model could generate images and video frames with pixel-level and accurate camera control and rebuild it in 3D. World Labs official presentation, where Atlas is pre-trained from scratch so that it can receive multi-modular input, including camera moves, and turn it into a 3D view. Atlas allows to use..

On September 2, Ali thousands asked today to announce the upgrade of the Qwen 3.8-Max model to version 0902. Officially, Qwen 3.8-Max was given further post-training around its programming (Coding) and professional office (Cowork) to enable the overall performance of the model to be broken down again, better suited to the real and complex tasks of the enterprise, scientific research, long-cycle tasks, etc. Qwen 3.8-Max rose 22 to 1691 in the Code Arena front-end programming master list. This means Qwen 3.8-M..

On August 29th, a new generation of flagship models was released and opened. The model uses a hybrid specialist structure with a total parameter of 770B, a activated parameter of 49B, a context length of 1M, and focuses on productivity enhancement tasks such as software engineering, game development, office and scientific research. 163 internal experts completed 203 project blinding, Hy4 preview average score 2.99 / 4.00; the government also demonstrated that the model was produced through MCP operation Unreal 5..

On August 26th, Microsoft CEO Satya Nadra, in his CNN dialogue, stated that businesses use AI not only to have “human capital” but also to accumulate “Token capital” made up of narratives, employee and model interactions, and how it is used. This knowledge should remain within the enterprise and not be tied to a single model or service provider. He argued that enterprises should retain the metadata generated by models every time they were used in order to train their weights or models; in the technical architecture, application layers should also be separated from models, and context and memory should be separated from models, so that..

ACCORDING TO THE 20 AUGUST MESSAGE, TANG JIE, THE CHIEF SCIENTIST OF THE BRAIN SPECTRA AI, GLM-5.3 WILL UPGRADE MODEL CAPABILITIES THROUGH EXPANDED TRAINING, RATHER THAN SIMPLY INCREASING THE NUMBER OF PARAMETERS. THE MODEL USES THE SAME BASE MODEL, STRUCTURE, TOTAL PARAMETER AND ACTIVATED PARAMETER LEVELS AS GM-5.2, BUT AFTER ONE MONTH OF LONG-CYCLE ENVIRONMENTAL TRAINING AND INTENSIVE LEARNING, THE MODEL PERFORMANCE HAS IMPROVED SIGNIFICANTLY. TANG JIE RECALLED IN HIS POST THAT EARLY STUDIES HAD SUGGESTED THAT MODELS SHOULD EXPAND THEIR PARAMETERS AT A FASTER RATE, BUT FOLLOW-UP STUDIES HAD SHOWN THAT TRAINING DATA NEEDED TO BE PRESERVED BETWEEN MODEL PARAMETERS AND TRAINING DATA IN THE CONTEXT OF LIMITED COMPUTING RESOURCES..

On August 15, Anthropic released his second Risk Report in the early hours of the morning, disclosing a model, Model 2, which exceeds Claude Mythos 5 and is not yet open to the public but has been used extensively within the company. According to the report, Model 2 is classified as "Mythos-class Model", which is stronger in some areas than Mythos 5 and has a slightly stronger overall performance. The model is now deeply involved in the research and development cycle of Anthropic

On August 8th, OpenAI released a statement that an internal assessment showed that Astra had made significant progress in the development of the model "Intelligent Encoding and Cybersecurity" and that, in conjunction with an expert assessment, the company "can't rule out" It has the potential to meet the "significant" Internet security threshold in the Preparadness Platform. OpenAI immediately suspended Astra ' s internal activities that did not meet the new security requirements and introduced stricter security controls for the higher-capacity model. By OpenA..

On 22 July, in an interview with the economic coverage of the twenty-first century, Huang Shui, the head of the dark side of the month, said that the value of the front model should not be measured only at low prices. This response was directed at the industrial controversy surrounding pricing, arithmetical tension and open source routes following the publication of Kimi K3. He indicated that Kimi would continue to be open-source, but open-source and closed-source services would be demand-driven: clients requiring privatization deployment and fine-tuning would choose open-source models, and clients wishing to purchase hosting services directly might choose closed-source services, which did not in themselves determine the quality and price of Token. It's called..

On July 20, the news that yesterday, Dean Ball, the head of OpenAI strategy, wrote that Kimi K3 was a "very good model" and that its performance could not simply be attributed to distillation. In the smart-body programming scenario, Kimi K3 has almost reached the level of the best open model in the first quarter of 2026; however, according to his experience of use, the model consumes more token and therefore the actual running costs are not necessarily low. Ball also stated that he was surprised that China was still allowed to open such high-capacity models. He thinks it's relevant..

On July 9th, SpaceXAI officially released Grok 4.5, which is positioned as the firm ' s strongest model to date, with the lead code, AI smart body mission and knowledge-based work. The model was combined with Cursor training based on a new V9 base, with parameters of approximately 1.5 trillion, using tens of thousands of British GB300 GPUs. In performance, Grok 4.5 outperformed Claude Opus 4.8..

According to information received on 2 July, according to WIRED, the United States Department of Commerce has recently withdrawn its export, re-export and internal transfer authorization requirements for the high-level models Fable 5 and Mythos 5 under the Anthropic flag. Anthropic official has also indicated on X that as of today, Claude Fable 5 visits will resume. It is worth noting that, on the eve of the lifting of the embargo, the Chinese group of developers is experiencing the harshest "flags" in history. According to Anthropic, transparency released in early 2026..

On July 1st, local time 30, Anthropic released a new generation of medium model Claude Sonnet 5, which strengthens the core of AI intelligence. Officially, the model ' s performance in self-planning, browser and end-of-service tools has reached the level of a previously large and expensive model. For performance, Sonnet 5 scored 63.2% in the programming baseline test, close to the top model Opus 4.8 and significantly superior to the previous Sonnet 4.6. Early partner feedback indicates that the model is self-..

On June 18, in a message issued yesterday (17 June) by Mask-based company XAI, it announced that it was officially on line with the Grok Imagine Video 1.5 model and ended with a preview of official access in xAI API. XAI launched the Grok Imagine Video 1.5 Preview preview on 3 June 2026, with the main hit "Step Chart One Generates Short Videos with Synchronized Audio" and, compared to the previous generation, the main upgrades to the synchronization, motor authenticity and production speed. I don't know

On June 16th, a Rio 3.5 Open 397B model under the umbrella of the municipality of Rio de Janeiro, Brazil, has recently been the subject of discussion in the AI community. The model page shows that it is open in the MIT protocol, with a total parameter of about 397B, single token activates about 17B and supports a longer context. However, the model was published less than 24 hours ago, and the open-source AI team, Nex-AGI, spoke on platform X, pointing out that Rio 3.5 was actually consolidated through the model (Model Mer..

In the news of 3 June, just now, Microsoft organized the Build 2026 Developer's Congress, which focused on the self-study model family, local AI development equipment, new AI terminal form and the corporate AI governance tool, announcing its comprehensive layout at the AI infrastructure level. Seven MAI models covering reasoning, codes, images, voice and transfer directions were introduced at a one-time event. MAI Thinking 1 uses a thin MoE structure, active 35B, total parameter size about 1T, supporting 256Kt..

Message from April 24, this morning, the DeepSeek-V4 model preview was officially online and synchronized with the open source. DeepSeek-V4 has a million-word super-long context, leading both domestic and open source areas in Agent capabilities, world knowledge and reasoning. The model is divided into two versions by size: the entry of the official network chat.deepseek.com or the official App will allow dialogue with the latest DeepSeek-V4 to explore the full new experience of 1M memory of the extra-long context. API Service Synchronized More..

In April 21st, yesterday, the dark side of the moon officially released and opened a new model, Kimi K2.6, which focused on upgrading the code, AI intelligence and office capacity. (a) Stronger code capability: the internal benchmark test score is about 201 TP3T, which can be coded continuously for 13 hours and handles over 4000 line codes to support multiple languages such as Rust, Go, Python

On April 16, according to The Information, Anthropic is preparing to launch the next generation of flagship models Claude Opus 4.7, and will simultaneously release an AI-driven design tool, both of which are the fastest available this week. The newly launched AI design tool is designed to help users to do ppt, website, drop-off page and product prototypes, directly against the goal ppt generation tool Gamma and Google Stitch. After the news came out, Adobe, Wix..

On April 8th, local time 7th, Anthropic released a new-generation flagship model, Claude Mythos Preview, and simultaneously announced the launch of Project Glasswing, which limited the authorization of the model to 12 technocrats, such as Apple, Microsoft, Amazon and others, for defensive cyber security, and was not open to the public. Mythos Preview shows a cliff-break lead in several benchmark tests. Anthropic says it's the biggest model it's ever trained, the power..

On March 19, in the early morning hours of this morning, Mimo V2 focused on three models: the flagship language base MiMo-V2-Pro, the full-mod Agent base MiMo-V2-Omni, and the voice synthesis model MiMo-V2-TTS. MiMo-V2-Pro: Total parameter mass exceeding 1T, activated parameter 42B, using the Hybrid Attention structure, with a mixed ratio of 5:1 in the previous generation to 7:1 and support for 1M 's hyper-long context..

On March 18th, OpenAI officially released today two new GPT-54 Mini and GPT-54 nano models, officially known as “the most capable mini-model to date”. The two models introduce the core GPT-54 competencies into a lighter structure, designed for high volume throughput and delayed-sensitive work loads. GPT-5.4 Mini has been significantly higher than GPT-5 Mini in code writing, reasoning, multi-modular understanding, and tool calls, and has been operating at more than two times faster. I don't know

On February 18, Anthropic released a new version of the size of the Sonnet model, continuing the company ' s four-month update. In the announcement of the new model, Anthropic highlighted its improvements in code capabilities, command compliance and computer operations. Sonet 4.6 will be the default model for free and Pro users. The sonet 4.6 test version will support the context window of 1 million tokens, which was previously the largest..

12 February, just now, the latest model GLM-5 is officially online and open. GLM-5 is described as a product of the move towards Agentic Engineering: in the Coding and Agent capabilities, its access to open source SOTA shows that it is approaching Claude Opus 4.5 in the use of a real programming landscape, with a special focus on complex system engineering and long-range Agent missions. GLM-5 uses a completely new base: the parameter size is expanded from 355B (activated 32B) to 744B..