A weekly AI industry report identified Claude Opus 5.5 as a newly launched Anthropic model and said its costs were reduced by 40%. The report did not specify whether this figure refers to API pricing, the cost of equivalent tasks, or internal inference costs, and provided no benchmark results, context specifications, or availability scope. It presents the cost change as a quantified detail of the model update, but exact pricing, capability changes, regional coverage, and availability timing should still be confirmed through Anthropic’s official announcement. The materials also provide no release date, user scope, or plan details.

A morning AI brief dated November 2 said Kuaishou subsidiary Kling AI had completed a capital increase totaling RMB 19.8 billion, with Kingdee, MetaX, and other institutions participating as new investors. The brief did not disclose the post-money valuation, each party’s subscription ratio, use of proceeds, or closing conditions, nor did it provide the investors’ original announcements. Information that can currently be confirmed is limited to the transaction’s completion, amount, and investor scope. As the valuation and ownership structure remain undisclosed, the company’s overall valuation or any individual investor’s stake cannot be inferred. The timing of fund transfers and specific future operating plans are still awaiting disclosure by the company.

Musk disclosed plans for Colossus 2 to add around 420,000 NVIDIA GB300 GPUs, with about 220,000 at Gun Range for training and around 200,000 at Spam for inference. The training cluster is planned to use an 800 Gbps high-speed network. Based on the plan, the report estimates that the total number of GB300s would reach around 860,000 after deployment, with a target of 1.2 million Blackwell GPUs by the end of 2026. The plans do not specify power supply, delivery, or commissioning timelines; the announced scale is a construction target, not a deployed result.

MiniMax announced on its official social media account that it had quantized M3.1 Flash, a model with 428 billion total parameters, to expand local deployment across different hardware configurations. The announcement mentioned the involvement of researchers and engineers from the Apple Silicon, AMD, and NVIDIA ecosystems. The company did not disclose the quantization bit width, minimum memory requirements, inference speed, changes in accuracy, or independent evaluation results, nor did it provide installation instructions or hardware requirements for each platform. The update broadens the scope for local deployment, but actual compatibility, deployment costs, and any quality loss still need to be verified through further configuration and testing.

A morning AI brief relayed a report from the CCID Research Institute saying that AI inference data volume had surpassed training data volume for the first time, with average daily Token calls exceeding 175 trillion. The report also said that large-model iteration cycles had shortened from several months to several weeks or even days. These figures reflect the growing share of inference services and product iteration in the AI industry, but the materials provide no report title, statistical methodology, sample scope, or cloud-provider data. The figures are currently only the brief’s account of the report and should not be treated as audited industry-wide statistics. The original report link was not included, and the complete data tables were not shown.

Cailian Press relayed findings from Hugging Face’s latest observational report, which said that in most months of 2026, the parameter-count ceiling for Chinese open-source models ranged from 754 billion to 2.78 trillion, while U.S. models were generally below 130 billion. More than 80% of models with over 20 billion parameters used permissive licenses, the report said. It also counted more than 150,000 models derived from Qwen, indicating a sizable ecosystem. The materials do not provide the full methodology for sample selection, parameter counting, or license classification. Cross-region comparisons should be understood according to the report’s definitions; monthly distributions and the complete data tables were also not shown.

An AI brief dated 10/31 relayed Cisco Talos researchers’ disclosure of the VOLTA information-stealing malware activity, saying it uses a machine-learning model to automatically assess and rank stolen corporate data. The materials provide no affected organizations, attack timeline, infection scale, or defensive measures, and do not explain how the model is integrated into the data-processing pipeline, so the scope of the activity and any losses remain unclear. The only information that can currently be confirmed is that the research team reported the use of machine learning to screen data after it was stolen. The brief also did not include a link to the original technical report.

Shanghai Jiao Tong University, Fudan University, and Tencent YouTu jointly introduced and open-sourced EvaOmni, a multimodal evaluation benchmark comprising 392 videos ranging from 1 to 12 minutes, 1,027 sets of human-annotated question-answer pairs, and 15 fine-grained capabilities. The benchmark requires models not only to answer questions but also to locate evidence segments in audio and video. The team tested 24 Omni LLMs, with Gemini 3.7 Flash achieving an answer accuracy of 79.4%. The design combines answer correctness with audiovisual evidence localization; the materials do not provide the full model rankings or reproduction details.

An arXiv paper uses the Joint-Embedding Predictive Architecture (JEPA) to analyze cross-modal representation alignment. It proposes that when a good view of the same scene is used as a positive sample, separately trained models for different modalities can learn representations that differ only by a rotation, and it provides an alignment error bound. The authors validate the scaling relationship using 145,000 captions and 940 million web images, claiming that for the same alignment performance, the number of good-view samples required is about 1/42 of the number required for next-token prediction. A text JEPA with 109M parameters outperformed DINOv2 and CLIP on alignment.