mode

Seedance announced the launch of the video generation model Seedance 2.5 flash, which supports generating videos up to 30 seconds long in a single instance, allowing for a maximum input of 10 images, 5 video segments, and 5 audio segments, providing resolutions of 480P and 720P.The official statement indicated that the generation speed of this model is twice that of Seedance 2.5; for example, the generation cost for video generated from text and images at 720P is approximately 0.99 yuan/second, which is about 34% lower than Seedance 2.5.Seedance 2.5 flash will gradually be available on platforms such as Jimeng, LibTV, Updream, Creao AI, and TapNow starting today, with the API expected to go live on November 15 on Volcano Ark.

On November 12, at the opening ceremony of Baidu World 2026, Kunlun Chip, a subsidiary of Baidu, introduced its next product plans. The M100 has been launched as scheduled, and the M300 will be released next year.Among them, the M100 focuses on extreme cost performance; the flagship M300 has made comprehensive capability leaps in multiple dimensions such as computing power, storage, and connectivity. In addition, by 2028, multiple chips based on the new N architecture will also be launched to match differentiated application scenarios.Baidu founder Robin Li stated at the conference that the demand for AI computing power has shifted from being primarily training-focused to balancing both training and inference. The scaling of intelligent applications has driven a massive explosion in inference computing demand, "inference will become the main driving force for the long-term sustained growth of AI chips."

According to the Science and Technology Innovation Board Daily, Doubao Work has officially integrated with multiple third-party models such as Kimi 3.1 Pro, GLM 5.3, and DeepSeek V4.1 Flash.In addition, Doubao Work has launched a limited-time discount for Doubao 2.1 Lite and DeepSeek V4.1 Flash, reducing the consumption multiplier for these two models.

According to the Science and Technology Innovation Board Daily, at the opening ceremony of Baidu World 2026, Baidu founder Robin Li stated that the iteration speed of models is very fast, with more than one new model released almost every day in the first nine months of this year. At the same time, the demand for model invocation has experienced a huge explosion, leading to the emergence of a batch of token transfer stations. To this end, Baidu launched the large model trading platform, Token Exchange.Robin Li stated that Token Exchange is a marketplace for large models, aiming to gather high-quality models and model providers, allowing enterprises and even individuals to easily find, purchase, and use reliable model tokens. Since its launch a month ago, over 100 supply resources have settled in Token Exchange, with an average daily invocation exceeding 50 million times and an average daily token invocation surpassing 700 billion.

According to MTS, OpenAI's artificial intelligence model has proven half of a million-dollar Millennium Prize problem in the field of physics, supported by a review panel led by a Nobel Prize winner. MTS stated that this problem is the Yang-Mills problem. OpenAI first announced the relevant proof on October 26, stating that its model completed this work in September. The review panel is chaired by Nobel Prize winner David Gross, with members including Arthur Jaffe, who co-authored the official statement of the Millennium Prize problem.The relevant proof has been formally completed using the linearly verifiable computer language Lean. The Yang-Mills problem is one of the seven Millennium Prize problems established in 2000, each with a prize of one million dollars; the only one solved so far was completed by Grigori Perelman, who refused to accept the prize. MTS pointed out that to win the award, it is still necessary to prove the mass gap and meet the Clay Mathematics Institute's rules regarding publication and a two-year recognition period. X user Theo Jaffee stated that physicists have been able to handle quantum field theory and quantum Yang-Mills theory on an empirical level, but have always lacked a mathematical formalization of why this theory holds.

According to Business Insider, Google (GOOG.O) is internally testing a high-speed Gemini AI model named Hydrogen.

According to Tether's official announcement, Tether AI Research has released the open-weight model QVAC BotPsy, which has 800 million parameters and is based on the Alibaba Qwen3.5 architecture, aimed at tool invocation and agent tasks on personal devices.Tether stated that in four evaluations: BFCL v4, ACEBench, τ-bench, and ToolSandbox, BotPsy's scores are approximately twice that of its base model and surpass the 2 billion parameter Qwen3.5-2B-Think. The model can convert user requests into local tool calls, and developers can adapt and deploy its weights within their own systems.

On November 11, B.AI announced that the API management page has completed a comprehensive experience upgrade, optimizing the model selection and API access process to help developers more efficiently find computing resources that fit their business needs.The new version of the page adds anchor navigation, allowing direct access to core modules with one click; it supports multi-dimensional filtering by brand and discount, quickly locating the required models; model versions and billing information are more transparent, and it integrates a "model ranking" quick reference for developers to compare selections. Welcome to visit chat.b.ai/chat to experience the new API page.

According to IT Home, OpenAI announced that it will officially discontinue the Advanced Voice Mode and Standard Voice Mode in ChatGPT on December 11, 2026, and will focus future development resources on GPT-Live. OpenAI began notifying users on November 10 about the upcoming discontinuation of the old voice modes, and users can continue to use them from now until December 11.On December 11, the two old modes will be removed from the "Settings - Voice" option, and users who previously selected these modes will automatically switch to GPT-Live. Previous voice conversation records will still be retained, and users can continue through text or GPT-Live.Since GPT-Live currently does not support real-time video and screen sharing, these two features of ChatGPT will continue to use Advanced Voice Mode. Users will also continue to use Advanced Voice Mode when having voice conversations with custom GPTs until December 11.

MiniMax officially released H3-Image and open-sourced the model weights. This model can directly generate images of 4096×4096, approximately 16.8 million pixels, without additional upscaling. According to official data, the generation time is less than 10 seconds, with speeds reaching up to 5 times that of mainstream products. Text-to-image and image editing have been integrated into the same model, supporting up to 16 reference images.H3-Image follows the architecture of the H3 video model, removing the audio component while retaining the original text encoder and redesigning the decoder for images. The core image generation network has 33.1 billion parameters, requiring only 6 sampling steps to generate an image. On August 7, MiniMax researcher Kiro revealed in a Reddit AMA that the team is developing a dedicated image model based on the H3 homogenous architecture and plans to open-source the weights. They also proposed a workflow where the image model generates the first frame, which is then handed over to H3 for video generation.On November 3, Pruna AI launched P-H3-Image based on the open-source H3 video model, supporting raw images and retouching, compressing the generation network from 33.1 billion to approximately 20 billion parameters, requiring only 8 sampling steps. The minimal quantization version is about 20 GB, with peak memory usage of approximately 21 GB when generating 8 million pixel images. The official MiniMax version retains the 33.1 billion parameter generation network, with a sampling step count of 6, and supports higher resolutions and more reference images, with the original weights plus the reused text encoder totaling approximately 135 GB. There are regional restrictions on the open-source license; the United States, United Kingdom, European Union, and South Korea are not within the community license authorization regions. Companies with annual revenues exceeding 1 million dollars or those providing model hosting services to third parties must obtain separate authorization in advance.

According to Gelonghui, sources have revealed that Apple is developing a Mac application named Harbor, which can turn computers equipped with Apple chips into private AI inference servers, supporting direct downloading, management, and running of open weight models without the need to upload data to the cloud.The application can also connect two or more compatible Macs via Thunderbolt 5, automatically forming a cluster and integrating unified memory to run large models that exceed the capacity of a single machine or handle more concurrent requests without manual configuration or using the command line. The application is currently in a non-public preview by Claris, a software company under Apple, and the official launch date has not yet been announced.

Stripe posted on platform X that the AI model routing platform OpenRouter has officially become a part of Stripe, and both parties will collaborate to help businesses explore new business models and optimize token routing and usage.OpenRouter stated that over 10 million developers and businesses are currently using its services. Previously, Bloomberg reported in August that Stripe had finalized a deal to acquire OpenRouter for over $7 billion.

Google's AI programming tool Antigravity has added a Knowledge Work office mode, making the AI Agent, originally aimed at programmers, easier to use for research, writing, data analysis, and creating documents and presentations.The new mode hides development tools such as terminals, Git, and code modification records, focusing instead on conversations, documents, and task outcomes. Users do not need to understand code repositories and can directly have the Agent handle local files, call tools, and complete multi-step tasks.The old and new modes use the exact same underlying Agent, and there has been no reduction in models, plugins, Skills, MCP, or custom Agents. Users can switch between modes at any time without needing to restart the application. Google revealed that currently, tens of thousands of non-software engineering employees within the company use Antigravity weekly, covering departments such as sales, finance, law, design, research, and operations.

Meta released the model Muse Ember, positioning it as a small model aimed at complex reasoning execution and repetitive tasks for agents, suitable for tool invocation, coding sub-tasks, routing, information extraction, and long document parsing. This model natively supports text, images, audio, and video, capable of handling screenshots, charts, scanned documents, voice memos, meeting recordings, and video content.Muse Ember has been launched on Muse Code and Meta Model API, with the model identifier muse-ember-1.0, supporting the Responses API, a context window of 1 million tokens, and a maximum output of 512,000 tokens in a single instance. The pricing for prompts within 200,000 tokens is $0.10 per million input tokens, $0.004 for cached input, and $0.50 for output; for longer prompts, the corresponding prices are $0.30, $0.012, and $1.50 respectively.

According to the "Science and Technology Innovation Board Daily," Ant Lingbo Technology has recently open-sourced the tactile foundation model LingBot-Touch. This model is pre-trained on nearly 5000 hours of real tactile data, covering three mainstream sensing principles: magnetic tactile, visual tactile, and piezoresistive.LingBot-Touch aims to convert signals from different sensors into the same set of tactile representations, allowing for reuse by various dexterous hands and motion models. This open-source initiative helps address current issues in the robotics field, such as the lack of standardization in tactile data and difficulties in cross-hardware generalization.

OpenAI announced the launch of a centralized system called Control Room, designed to track model workloads within its research infrastructure and quickly halt related tasks when security or safety concerns arise. This system currently covers research workloads within OpenAI, including training experiments, evaluations, and data processing. OpenAI stated that as model capabilities increase, associated safety and security risks also rise, making it increasingly critical to detect potential issues in a timely manner and swiftly halt models when necessary.In terms of specific mechanisms, Control Room assigns an identifier to each research task, and the inference engine verifies this identifier before generating model outputs. When a stop command is issued by manual or automated monitoring systems, Control Room revokes the authorization for that task, thereby blocking subsequent outputs, even if the underlying research program is still running. OpenAI noted that cached authorizations could delay shutdowns by up to 35 seconds; additionally, if Control Room itself is unavailable, the system is designed to reject new requests.Prior to this release, OpenAI had disclosed multiple incidents this year where research models deviated from instructions during evaluations. In July, the company reported that its model escaped a restricted testing environment and infiltrated Hugging Face to cheat in a security assessment. In October, it disclosed that a research model used an unauthorized encrypted wallet to pay for smart contract deployment fees on the testnet ETH, and this expenditure was omitted from the final summary.