Huawei Cloud has launched the open-source openPangu-2.1-Pro model on its MaaS platform, with updates covering Agent development, AI coding and inference optimization. Its Agent capabilities were optimized using more than 100,000 specialized data entries, and integration with several Huawei Cloud products is planned.
Huawei Cloud has made the open-source Pangu model openPangu-2.1-Pro available on its Model as a Service (MaaS) platform; the source does not specify the launch date. The update covers Agent development, AI coding and inference optimization. According to Huawei Cloud, more than 100,000 specialized data entries were used to optimize Skills triggering, tool selection and tool invocation, covering scenarios such as deep research and HarmonyOS system operation. Coding optimizations target HarmonyOS App development and Ascend operator generation to improve the accuracy of ArkTS code generation. Huawei Cloud also says that optimizing key operators and inference paths has improved throughput efficiency in typical deployment scenarios, but the source provides no specific improvement figure. The model will be integrated into Huawei Cloud products including Zhigu AgentArts, Madao CodeArts and Guoban OfficeAce.
Agnes AI open-sources 187.8B-parameter multimodal MoE model
模型发布
Agnes AI has released Agnes 2.5 Pro Theta, a multimodal MoE model with approximately 187.8B parameters and BF16 weights for local inference. Its configured context limit is 262,144 tokens, while the unquantized weights alone occupy approximately 349.7 GiB.
Agnes AI has released Agnes 2.5 Pro Theta under the Apache License 2.0, providing BF16 weights for local inference through Transformers. The package includes the model implementation, tokenizer, chat template, and image and video processing components, with approximately 187.8B total parameters. The publisher-provided evaluation snapshot covers eight benchmarks, reporting 92.9% on GPQA Diamond, 90.5% on LiveCodeBench v6, and 84.7% on AA-LCR; however, it does not include exact prompts, sampling settings, token budgets, Agent scaffolds, or raw run logs. The configured context limit of 262,144 tokens does not mean that every length or modality combination has been tested. The unquantized weights alone occupy approximately 349.7 GiB, with additional memory required for loading, activations, caches, and visual inputs. According to the publisher, the checks performed cover only model construction and input preparation, rather than full end-to-end generation validation; the reported scores also do not establish comprehensive image and video performance or long-context reliability.
OpenCode Go subscriptions now support payments via 微信支付
开发生态
OpenCode has added WeChat Pay for Go and Go Plus, making the payment option available to both new and existing subscribers. A one-time authorization in WeChat enables monthly automatic renewals; Teams is not supported, and billing remains in US dollars.
OpenCode has announced in its Console changelog that Go and Go Plus subscriptions now support WeChat Pay. The rollout does not include Teams subscriptions, and billing remains in US dollars. New users can select WeChat Pay in the Go tab of Console to subscribe, while existing subscribers can switch their payment method to WeChat Pay on the Manage subscription page. According to the official explanation, a one-time authorization in WeChat allows the subscription to renew automatically each month. If a user cancels the payment authorization in WeChat, confirmation will be required again at the next renewal.
OpenAI launches science blog and publishes an analysis of Standard Model research
技术与洞察
OpenAI has launched a science blog with an article explaining Nathaniel Craig’s work with GPT-6 Astra on the Standard Model. The paper and verification materials were already public, so the post is not the research’s initial release.
OpenAI announced the launch of the OpenAI science blog, publishing its first guest post by Nathaniel Craig at launch to explain previously released research on the Standard Model. Craig is a visiting scientist at OpenAI and heads the particle theory group at the University of California, Santa Barbara. According to the article, he worked with GPT-6 Astra to propose a candidate nonperturbative definition of the Standard Model on a lattice, with the model leading the completion of the detailed proof. The associated preprint, mathematical supplementary materials and Lean formalization archive were already available on arXiv and Zenodo; the blog post explains the research to a broader readership. OpenAI says the new blog will share findings, practical methods and researchers’ firsthand experiences from efforts by OpenAI, its collaborators and the scientific community to advance scientific discovery using AI.
Jensen Huang reportedly considering a return to NVIDIA as CEO
前瞻与传闻
Jensen Huang has discussed a possible return to NVIDIA with close confidants, according to the Financial Times, while allies are lobbying the board to reinstate him as CEO. A petition supporting his reinstatement had gathered 156,850 signatures at publication, but his return remains uncertain.
Former NVIDIA CEO Jensen Huang had discussed a possible comeback with close confidants by the time of publication, according to the Financial Times, while his Silicon Valley allies were organizing a lobbying effort to seek his reinstatement from the board. Whether he will return, and in what role, remains undecided. On the shareholder side, Altimeter Capital CEO Brad Gerstner and Atreides Management’s Gavin Baker are leading pressure on the board to keep a return open to Huang. Altimeter Capital holds approximately 9.4 million NVIDIA shares worth nearly $1.9 billion, while Atreides Management holds approximately 1.2 million shares.
The petition on the “I Stand With Jensen” website calling for his reinstatement has collected 156,850 signatures, including more than 7,500 current NVIDIA employees. Signatories also include Elon Musk, Replit CEO Amjad Masad, Gavin Baker and Tim Sweeney. Meanwhile, Huang remains under an internal NVIDIA investigation into his alleged relationship with Alice Huang. U.S. Attorney General Todd Blanche said the Department of Justice had not reached a conclusion on the espionage allegations against Alice Huang. Singapore’s Ministry of Home Affairs formally denied a Wall Street Journal report claiming that local officials were concerned about her close relationship with Jensen Huang and had investigated her as a suspected Chinese spy.