月之暗面 Releases Kimi K3 Model Weights and Full Technical Report
要闻
月之暗面 officially released the Kimi K3 model weights and published the full technical report. Kimi K3 has a total of 2.8 trillion parameters, with approximately 104 billion parameters activated per token. The technical report systematically introduces the core design of the model. Additionally, 月之暗面 open-sourced three key infrastructure technologies: MoonEP, FlashKDA, and AgentEnv. The model is released under the Kimi K3 License, which includes restrictions such as MaaS providers that reach a certain revenue scale needing to sign an agreement with 月之暗面.
Moonshot AI officially released the Kimi K3 model weights and the full technical report. Kimi K3 adopts a Mixture-of-Experts (MoE) architecture with a total parameter scale of 2.8 trillion, activating approximately 104 billion parameters per token, and setting up 896 routed experts, from which 16 are selected for computation. The model also possesses native visual understanding capabilities and supports a context length of up to 1 million tokens.
In the technical report, Moonshot AI detailed several key designs. Among them, Kimi Delta Attention and Gated MLA are combined proportionally to reduce the computational cost of long contexts; Attention Residuals are used to improve cross-layer information transmission; Stable LatentMoE enhances the training stability of ultra-high-sparsity models through new activation functions and load-balancing methods; and MoonViT-V2 trains the vision encoder using next-token prediction from scratch.
In addition to the model weights and technical report, Moonshot AI also simultaneously open-sourced MoonEP, FlashKDA, and AgentEnv. These three technologies are respectively oriented toward large-scale MoE expert-parallel communication, high-performance KDA operators, and distributed agent training environments, covering important infrastructure components of Kimi K3 from pre-training to reinforcement-learning post-training.
The code and weights of Kimi K3 are released under the Kimi K3 License, which includes certain restrictions. For example, MaaS enterprises that reach a certain revenue scale are required to sign a separate agreement with Moonshot AI before using the model and its derivative versions for commercial purposes.
Tibo, head of OpenAI Codex, hinted that usage limits would be reset within hours to celebrate the rapid adoption of ChatGPT Work. Meanwhile, Greg Brockman announced that ChatGPT Enterprise customers who register before August 21 will receive up to $200 in usage credits valid for 14 days per team member within two weeks when they first try ChatGPT Work.
OpenAI Codex lead Tibo recently posted, saying that they are celebrating the rapid adoption of ChatGPT Work and feel that a "limit reset" is coming. He also reminded users to "hold on tight to your ultra and /fast." However, Tibo did not specify the exact timing and scope of the reset, only stating that he would "be back at the computer in a few hours." Meanwhile, OpenAI co-founder Greg Brockman announced on social media that ChatGPT Enterprise customers who register before August 21 will receive up to $200 in usage credits for each team member's first trial of ChatGPT Work, valid for two weeks, with a validity period of 14 days.
微信WeLM Introduces HD4-80B and 617B-Parameter MoE Models
模型发布
The 微信WeLM technical team released large-scale MoE models WeLM-HD4-80B and WeLM-HD4-617B using Hidden Decoding technology. According to the official, both models have improved in complex reasoning, code generation, and general conversational ability.
The WeLM technical team at WeChat developed MoE models with 80B and 617B parameters using Hidden Decoding technology. The team claims improvements in capabilities such as complex reasoning, code generation, and general conversation, aiming to serve WeChat's massive user base. The team has not released details on open downloading or API usage of the models.
微软 Unveils First Cybersecurity Model MAI-Cyber-1-Flash
模型发布
微软 announced its first cybersecurity model, MAI-Cyber-1-Flash, which is deeply integrated into the multi-agent vulnerability detection and remediation tool MDASH. According to the official, the combination scored 96% on the CyberGym security benchmark, with costs reduced by 50% compared to the previous best solution.
Microsoft CEO Satya Nadella announced a series of security updates, centered on MAI-Cyber-1-Flash—Microsoft's first cybersecurity model, built specifically to discover hard-to-find vulnerabilities in complex codebases, and deeply integrated into the MDASH multi-agent security tool. According to official data, the combination scores 96% on the CyberGym benchmark, 12 percentage points higher than Mythos, surpassing Gemini and GPT. MAI-Cyber-1-Flash handles about 90% of tasks, while complex tasks are delegated to GPT-5.4, reducing overall costs by 50% compared to previous MDASH best solutions. Microsoft also launched the agentic security system Project Perception, providing agent teams for continuous monitoring, patching, and closing threat vectors.
Qoder Upgrades '极致免费用' Campaign, Extra Pack Resets Daily for 5 Days
开发生态
Qoder announced that the "极致免费用" campaign has been extended to July 31. From July 27 to 31, the 1,000-use "极致" bonus pack resets daily at 14:00, and the expert team mode fully supports free deduction. Everyone can claim a 200-use trial pack without threshold, while the 1,000-use bonus pack is automatically issued to historical subscribers or targeted at active users.
Qoder has extended the "Extreme Free" event to July 31 and upgraded daily benefits. During July 27-31, the 1,000-query "Extreme" bonus pack will automatically reset every day at 14:00 (UTC+8), for five consecutive days. Additionally, the expert team mode now supports deduction by free credits. The 200-query trial pack is available to everyone without restrictions, while the 1,000-query bonus pack is automatically issued to historical subscribers or targeted active users.
OpenRouter Teams Up with OpenAI to Offer Half-Price Deals on GPT-5.6 Terra and Luna
开发生态
OpenRouter is currently partnering with OpenAI to offer half-price promotions on GPT-5.6 Terra and Luna for a limited time.
OpenRouter officially announced that the GPT-5.6 Terra and Luna models currently enjoy an exclusive 50% discount on its platform. This promotion is a limited-time campaign in collaboration with OpenAI. The discount applies only to first-party OpenAI providers and covers input, output, and cache prices for the models. In addition, the Pro versions of these two models will also receive the same 50% discount.
OpenAI GPT-Live Voice Feature Opens to Education and Enterprise Subscription Plans
产品应用
OpenAI has opened the GPT-Live 1 voice feature to users on Education, Business, and Enterprise subscription plans. The feature provides a voice interaction experience close to natural conversation.
OpenAI announced that the new-generation voice model GPT-Live 1 is now globally available to ChatGPT's Education, Business, and Enterprise subscription plans. The feature had already started rolling out previously, and this expansion broadens its availability to offer more natural human-machine voice interaction. Users can experience it directly in ChatGPT's voice mode.
美团 Officially Launches Full-Scenario AI Agent Platform CatPaw
产品应用
美团 officially launched the full-scenario AI Agent platform CatPaw. The platform offers a multi-device synchronized intelligent workbench and enterprise-grade Managed Agents hosting capabilities, with two plans for individual users: Experience and Professional.
Meituan officially launched CatPaw, an all-scenario AI Agent platform, offering intelligent workbenches on desktop, mobile, and cloud, as well as enterprise-level Managed Agents development and hosting capabilities. The platform supports real-time synchronization across macOS, Windows, iOS, and Android, with cloud mode running 7x24 hours. It integrates Meituan's local-life industry insights. CatPaw supports autonomous collaboration among multiple Agents, persistent memory across conversations, and built-in expert and skill libraries. It has already covered 90,000 employees within Meituan and built 30,000 Agents. For individual users, it offers two plans: an Experience edition and a Professional edition. The Professional edition includes 4,000 Credits per month, with additional credit packs available for purchase. Meituan is now inviting its partner merchants to try it out first.
千问办公 Officially Launched, Integrated with DingTalk Ecosystem and Supporting Multi-Device Collaboration
产品应用
阿里巴巴 officially launched the AI office platform 千问办公. The platform can directly generate office deliverables such as PPT through natural language conversations. It now fully covers desktop, web, and 钉钉, and offers a free version as well as a subscription plan starting at 78 yuan per month.
Alibaba launched Qianwen Office, a one-stop AI office platform that allows users to directly execute complex tasks such as data analysis, PPT generation, and video editing through natural language conversations, and obtain usable results. The platform is based on the Qianwen series of large models, supports multimodal understanding, and can generate outputs in various formats such as PPTX, Word, Excel, and HTML. Qianwen Office now fully covers desktop and web endpoints and is deeply integrated with the DingTalk ecosystem. It offers a free version, a Personal Standard edition starting at 78 yuan/month with continuous subscription, and a Personal Premium edition starting at 158 yuan/month. The free version's daily login credits and new-user credits are limited-time activities, and certain capabilities such as browser automation are currently supported only on desktop.
腾讯QQ Announces QQ Pet Upgraded Return: Integrated with Hy3 Large Model for Proactive Responses
产品应用
腾讯QQ announced the full upgrade and return of QQ宠物. The new QQ宠物 features a 3D design and integrates 腾讯's Hy3 large model, allowing it to proactively respond to users. Users can now search for it in mobile QQ to claim it.
Tencent QQ announced the upgraded return of QQ Pet, now available on mobile QQ. The new QQ Pet features a 3D avatar and is integrated with Tencent's Hy3 large model, enabling the pet to proactively respond and offering multiple selectable personalities. The product retains classic gameplay such as feeding and bathing, and adds social features including the pet appearing in multi-screen scenarios like the chat list and friends visiting and stepping on each other. Currently, users can search for "QQ Pet" on mobile QQ to claim it.
NVIDIA Establishes Open Secure AI Alliance with More Than 30 Companies
行业动态
NVIDIA, together with more than 30 companies, established the Open Secure AI Alliance to develop and share open AI security defense tools and technologies. Currently, alliance members have contributed multiple open-source projects.
NVIDIA, together with more than thirty enterprises and institutions including Hugging Face, Microsoft, IBM, Red Hat, and SpaceXAI, has jointly established the Open Secure AI Alliance, dedicated to developing and sharing open technologies, tools, and methods to secure software and intelligent agents in the AI era. The alliance is built upon the Linux Foundation's Akrites initiative and OpenSSF community work, advocating that open models and open frameworks are critical to cybersecurity defense because they democratize defensive capabilities, enhance transparency, and allow defenders to run cutting-edge AI on their own infrastructure. NVIDIA has released the NOOA open-source framework on GitHub, Hugging Face has contributed the Safetensors secure weight format to the PyTorch Foundation, and SpaceXAI has open-sourced Grok Build, a terminal programming agent, with plans to open-source the weights of the Grok series models. The alliance also calls on policymakers to view open models as defensive assets rather than security risks.
SSI Announces Long-Term Strategic Partnership with NVIDIA
行业动态
SSI announced a long-term strategic partnership with NVIDIA. NVIDIA will make a significant investment in SSI, increasing SSI's computing power tenfold within the next 12 months. SSI stated that its research has reached a stage worthy of scaling.
SSI officially announced a long-term strategic partnership with NVIDIA, with NVIDIA making a significant investment in SSI. SSI stated that this collaboration will increase its computing power by 10 times within the next 12 months, adding that "our research has reached a stage worthy of scaling." NVIDIA will provide SSI with access to its next-generation Vera Rubin computing platform. WSJ reported that NVIDIA made the investment decision after obtaining rare access to SSI's confidential research, but the specific amount of the deal was not disclosed.
Relevant Authorities Respond to U.S. Plan to Sanction Chinese AI Companies: Oppose Smears and Sanction Threats
行业动态
In response to the U.S. plan to investigate Chinese AI companies for "distilling" U.S. models, the relevant authorities expressed strong opposition, stating that the accusations lack factual basis, and urged the U.S. to stop smearing and sanction threats.
In response to public statements by some senior U.S. officials expressing intentions to investigate and impose sanctions on Chinese AI companies over the alleged "distillation" of American frontier models, relevant authorities have responded. China has consistently opposed the politicization of technology and trade issues by the United States, pointing out that such accusations ignore the fact that some Chinese models are already at the forefront, lacking factual basis and legal justification. It is understood that many U.S. AI companies have distilled Chinese models during R&D and training, and numerous American enterprises oppose cutting off access to Chinese open-source models. China urges the United States to listen to objective and rational voices from the industry, stop defamation and sanction threats, and emphasizes that it will take all necessary measures to protect its interests against any actions that substantially harm China's interests.
Anthropic Clarifies Position on Open-Weight Models
行业动态
Anthropic CEO Dario Amodei stated that Anthropic has never advocated banning open-weight models. He suggested that AI safety risks should be addressed by restricting chip exports, combating industrial-scale distillation, and mandating safety testing, rather than banning open-weight models themselves.
Anthropic CEO Dario Amodei stated on the official blog that Anthropic has never advocated for banning open-weight models, calling the allegations false. Recent reports indicate that some U.S. officials are considering prohibiting American companies from using Chinese open-weight models, and multiple tech companies have signed an open letter in support of open weights. Amodei believes that protectionist bans cannot address the most critical national security concerns. He genuinely supports three measures: restricting the export of high-performance chips and chip manufacturing equipment to China, cracking down on industrial-scale distillation operations, and requiring all sufficiently powerful models, whether open or closed, to undergo mandatory safety testing. He partially agrees with the open letter's view that open weights promote competition, but disagrees with the assertion that open weights inherently make safety protections easier.
Alexandr Wang Confirms Meta Will Release New Model and Harness
前瞻与传闻
Alexandr Wang confirmed that Meta will release a harness and open-source models. However, the specific release date and model scale have not been announced.
RK posted on social media, citing Alexandr Wang's remarks, that Meta will launch a harness and open-source models, and Alexandr Wang subsequently replied to confirm. Social media users added that Meta signed the Open AI Alliance agreement earlier. The specific release time, model scale, and availability scope of the harness and open-source models have not yet been announced.
Nvidia and OpenAI Reportedly in Talks Over Data Center Financing Guarantee
前瞻与传闻
According to reports, Nvidia is in talks with OpenAI to provide approximately $250 billion in financing guarantees to support OpenAI's lease of a 10-gigawatt data center project developed by SoftBank in southern Ohio. The negotiations are still in early stages and may fall through or have adjusted terms.
According to a report by The Wall Street Journal, citing sources familiar with the matter, Nvidia is in talks with OpenAI to provide approximately $250 billion in financing guarantees to help OpenAI lease a 10-gigawatt data center project developed by SoftBank's energy subsidiary in southern Ohio. The total cost of the project is expected to exceed $500 billion, including chip costs, making it the largest data center project announced to date. Sources familiar with the matter said the guarantees cover data center leasing and debt financing, but do not include Nvidia's chip costs. Nvidia is also in discussions to provide OpenAI with up to $350 billion in chip procurement financing. Reuters said it could not independently verify the report, and the companies involved have not yet responded; Bloomberg, citing sources, said the talks are still in the early stages and could fall through or the financing terms could change.
Note: This content was created with AI assistance and may contain hallucinations and errors.
