Moonshot Releases 2.8T Kimi K3 Open Weights as Industry Debates Security and Regulation

Moonshot Releases 2.8T Kimi K3 Open Weights as Industry Debates Security and Regulation Moonshot AI has published the weights for its 2.8-trillion-parameter Kimi K3 model, while Anthropic addressed its stance on open weights and Microsoft unveiled its MAI-Cyber-1-Flash security model. Moonshot AI published the weights and architecture for Kimi K3, a 2.8-trillion-parameter system that rivals top frontier models. Anthropic CEO Dario Amodei stated his company does not advocate banning open-weight models, urging focus instead on chip controls and distillation enforcement. Microsoft introduced MAI-Cyber-1-Flash, a bug-hunting cybersecurity model embedded in MDASH that scored 96% on CyberGym. Safe Superintelligence partnered with Nvidia via a reported $5B investment, while Nvidia is in talks for a $250B financing guarantee for an OpenAI datacenter. Moonshot Releases 2.8T Kimi K3 Weights Moonshot AI released the weights and technical documentation for Kimi K3, creating the largest open-weight model to date at 2.8 trillion parameters. Operating the model requires high-end hardware, and Moonshot has also licensed out supporting infrastructure like attention kernels and agent systems. The release follows U.S. allegations that Moonshot distilled American AI models to build K3, an accusation Beijing rejected as AI hegemony. Anthropic Clarifies Stance on Open Weights Anthropic CEO Dario Amodei clarified the company's position on open-weight AI after abstaining from an Nvidia-led coalition letter signed by 50 companies, including OpenAI and Google. Amodei stated that Anthropic has never sought to ban open-weight systems, but argued against the notion that open models inherently strengthen security defense. He urged regulators to focus primarily on chip export controls, distillation crackdowns, and broader safety evaluations. Microsoft Debuts MAI-Cyber-1-Flash Integrated into Microsoft's MDASH agent framework, MAI-Cyber-1-Flash is designed to find and patch software vulnerabilities at half the cost of top competitors. The model reached 96% on the CyberGym codebase security benchmark, outscoring Anthropic's Mythos by 12 points. Microsoft also announced Project Perception, which deploys multi-agent groups to independently simulate attacks, investigate threats, and remediate code. Quick Industry Developments Ilya Sutskever's Safe Superintelligence secured an Nvidia partnership that reportedly includes a $5 billion investment. In addition, Nvidia is reportedly in discussions to guarantee $250 billion in financing for an OpenAI 10GW datacenter campus in Ohio. Anthropic clarified that exposed Claude conversations appearing in search results stemmed from user-shared public links, and Cursor introduced a localized $7 monthly tier in India.

月之暗面开源史上最大模型Kimi K3，Anthropic与微软同步披露最新AI动作 本期动态聚焦月之暗面正式开源2.8万亿参数的Kimi K3模型，同时涵盖Anthropic对开放权重监管的立场表态、微软推出的低成本网络安全模型，以及多家前沿AI机构的融资与业务进展。 月之暗面开源拥有2.8万亿参数的Kimi K3权重及核心技术栈，成为史上最大开源权重模型。 Anthropic首席执行官澄清从未主张封杀开源模型，呼吁重点关注芯片管制与模型蒸馏审查。 微软推出首款原生集成于MDASH的网络安全模型MAI-Cyber-1-Flash，宣称降低50%使用成本。 SSI获英伟达传闻达50亿美元的重大投资；英伟达拟为OpenAI数据中心提供2500亿美元融资担保。 月之暗面发布史上最大开源模型Kimi K3 月之暗面正式开放了Kimi K3的模型权重和技术报告。该模型拥有2.8万亿参数，在各项基准测试中对比Claude Fable 5与GPT-5.6-Sol，成为迄今为止最大的开源权重模型。月之暗面同时开源了注意力核心与智能体架构等底层技术，允许第三方托管并商用。对于此前美方官员指责其通过蒸馏美国模型构建K3，北京方面反驳称这是“典型的AI霸权行径”。 Anthropic阐明开放权重立场 Anthropic首席执行官Dario Amodei发文回应开源模型争议，指出公司从未提倡禁止开放权重模型。Amodei表示虽然不认同开源权重必然增强网络防御能力的观点，但主张将监管重心放在芯片出口管制、打击模型蒸馏以及扩大安全测试上，而非限制权重本身。与此同时，由英伟达牵头、OpenAI与谷歌等加入的开放权重支持信签名方已增至50家。 微软推出代码安全模型MAI-Cyber-1-Flash 微软推出专用于寻找软件漏洞的代码安全模型MAI-Cyber-1-Flash，直接整合进MDASH智能体系统。该模型在CyberGym代码库安全基准测试中达到96%的准确率，较Anthropic Mythos高出12点，且宣称成本降低50%。此外，微软还发布了能够自主模拟网络攻击、调查威胁并修复漏洞的Project Perception多智能体系统。 行业其他重要速讯 Ilya Sutskever创立的Safe Superintelligence（SSI）宣布与英伟达建立合作，据悉获得英伟达约50亿美元投资；英伟达正商讨为OpenAI在俄亥俄州的10GW数据中心租赁提供2500亿美元融资担保。此外，部分Claude用户对话因链接公开被搜索引擎收录；Cursor在印度推出7美元/月的Start计划；Sam Altman则在播客中公开宣称技术已踏入奇点阶段。

[Read Online]({{live_url}}) ^| [Sign Up](https://www.therundown.ai/subscribe?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=moonshot-lets-history-s-largest-open-model-loose) ^| [Advertise](https://therundownai.typeform.com/to/kraZ1TSO?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=moonshot-lets-history-s-largest-open-model-loose) **Good morning, {{ first\_name | AI enthusiasts }}, and welcome to the 5,588 new readers who joined us yesterday.** The U.S. is still weighing what to do about powerful open models, and (most of) the industry has lined up on the open side. Both just got handed a live test case. Moonshot just delivered on its promise to publish the weights behind its near-frontier system, making K3 the biggest open model ever released — and likely the most consequential one, too. *P.S. — Something new just landed: The Rundown's AI Workflow Hub is officially open. Share what you're building, see what others have made, or team up on something new. Post something great, and you might just see it featured right here in the newsletter. Come see what we're building.* **In today’s AI rundown:** Moonshot publishes weights for powerful Kimi K3 Anthropic sets the record straight on open weights Make precise edits on real product photos with AI Microsoft’s powerful new cybersecurity model 4 new AI tools, community workflows, and more **LATEST DEVELOPMENTS** Image source: Moonshot AI **The Rundown:** Moonshot AI just officially [released](https://huggingface.co/moonshotai/Kimi-K3?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=moonshot-lets-history-s-largest-open-model-loose) the weights and tech report behind its powerful Kimi K3, letting anyone (with serious GPUs) download and run the 2.8T-parameter model, the largest ever put out with open weights. **The details:** K3 was released earlier this month, making waves after competing with both Claude Fable 5 and GPT-5.6-Sol across a variety of benchmarks. Free weights don’t mean free to run, with the large model requiring much heavier GPUs than the average consumer hardware to run locally. Moonshot also opened core pieces of its stack, like attention kernels and agent infrastructure, under a license letting anyone host and sell K3 access. U.S. officials [accused](https://www.therundown.ai/p/openai-cyber-test-escapes-the-lab?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=moonshot-lets-history-s-largest-open-model-loose) Moonshot of distilling U.S. models (copying their outputs) to build K3, which Beijing [rejected](https://www.reuters.com/world/china/china-accuses-us-ai-hegemonism-threatens-countermeasures-over-potential-probes-2026-07-27/?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=moonshot-lets-history-s-largest-open-model-loose) as "typical acts of AI hegemony." **Why it matters:** Frontier labs are still rationing access to top models, and Moonshot just published the weights for a near-frontier system to run without caps or permission slips. While the U.S. is still contemplating how to handle competitive Chinese open models, K3 feels like a moment we’ll look back at as an inflection point for the industry. **The Rundown:** If your team still sorts through customer feedback manually (or with a patchy mix of manual work and AI), there's a better way. Unwrap brings all your feedback across surveys, reviews, tickets, social comments etc. into one view, then uses AI and NLP to surface what actually needs your attention. **With Unwrap, you can:** Automatically categorize every piece of customer feedback Query feedback via Unwrap Assistant, in Slack, or through Unwrap's MCP Get real-time alerts and a clear view of customer sentiment Use a platform trusted by companies like Clay, Perplexity, Oura, Stripe, and Lululemon Unwrap is offering a trial of its tools to Rundown AI subscribers! **[Just grab 20 minutes with the team to get set up](https://cal.com/unwrap/quick-connect?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=moonshot-lets-history-s-largest-open-model-loose)**. Image source: Anthropic **The Rundown:** Anthropic CEO Dario Amodei [published](https://www.anthropic.com/news/position-open-weights-models?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=moonshot-lets-history-s-largest-open-model-loose) a post laying out where the company stands on open-weight models, coming as the company remains a key signatory [holdout](https://www.rundown.ai/articles/anthropic-opus-5-surprise?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=moonshot-lets-history-s-largest-open-model-loose) in the "Open Weights and American AI Leadership” [letter](https://images.nvidia.com/pdf/Open-Weights-and-American-AI-Leadership.pdf?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=moonshot-lets-history-s-largest-open-model-loose). Amodei said, “Anthropic has never advocated for a ban on open-weights models”, even though it would “protect US AI companies from competition”. The CEO instead pushed support for chip controls, distillation crackdowns, and broader safety testing, issues that Anthropic has “consistently advocated for”. The Nvidia-led letter defending open models has grown to 50 signers from the [initial 25](https://x.com/jensenhuang/status/2080643682408321103?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=moonshot-lets-history-s-largest-open-model-loose), with OpenAI and Google as notable additions. Amodei also said he agrees with “much of” the letter, but doesn’t believe open-weights models lead to better security safeguards and help defenders. **Why it matters:** With K3 nearing the frontier and rumblings of U.S. bans, the open-weights fight is a major conversation that has been happening very loudly without Anthropic. Amodei’s argument is that the weights aren’t the threat; it’s the chip access and distillation that risk authoritarian regimes gaining the AI power he’s worried about. **The Rundown:** In this guide, you’ll learn how to use Reve 2.1 to change the color of one small part of a product photo without regenerating the entire image. You’ll create a new copy with changes while keeping the rest of the original image intact. **Step-by-step:** [Sign up for a free Reve account](https://reve.com/?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=moonshot-lets-history-s-largest-open-model-loose), choose Upload and edit, and upload the product image(s) you want to change Reve will automatically scan your image and break it into editable parts. You’ll see dropdown menus for each part that get more detailed as you open them In that parts dropdown menu, click the color palette tool, then paste in the new color hex code you want to test. Repeat this for any other parts if needed When you’re done picking colors, click the Apply button. Reve will duplicate the original image and create a new version with all of your edits. **Pro tip:** You can make other edits with the same workflow, such as changing a shadow, material, or background. [Get resources, prompts, and a full step-by-step video guide here.](https://app.therundown.ai/guides/make-precise-image-edits-on-real-photos-with-ai?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=moonshot-lets-history-s-largest-open-model-loose) **The Rundown**: Outreach, the leading agentic AI revenue platform, explores how modern revenue teams are cutting through signal overload – using AI to prioritize what matters most and turn scattered intelligence into consistent, high-quality execution. **In the on-demand session, you'll learn:** How top teams are shifting from volume to precision Where AI improves prioritization and decision quality How better focus translates to more pipeline Image source: Microsoft **The Rundown:** Microsoft just [launched](https://microsoft.ai/news/introducing-mai-cyber-1-flash-inside-mdash/?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=moonshot-lets-history-s-largest-open-model-loose) MAI-Cyber-1-Flash, its first cybersecurity model integrated directly into the company's MDASH agent system to hunt software bugs, promising what CEO Satya Nadella calls "frontier-grade security at half the cost." Paired with MDASH, the model hit 96% on the CyberGym benchmark for security on large codebases, coming in 12 points above Anthropic's Mythos. Cyber-1-Flash claims a 50% cost reduction over top models, with the efficient system able to handle 90% of tasks in workflows over more expensive rivals. Microsoft also introduced Project Perception, a tool that uses teams of agents that simulate attacks, investigate threats, and repair flaws on their own. Microsoft AI CEO Mustafa Suleyman said, "token cost is now the real constraint for defenders," calling for "continuous, real-time, cost-efficient agents." **Why it matters:** Coming just a week after the OAI / Hugging Face incident that was the talk of the AI security world, Microsoft’s new model is a timely release and possibly one of the first in a wave of focused releases that will accelerate both vulnerability detection and remediation for enterprises. **QUICK HITS** 🎯 [Kimi K3](https://www.rundown.ai/tools/kimi-k3?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=moonshot-lets-history-s-largest-open-model-loose) - Moonshot’s powerful open-source model 🔒 [MAI-Cyber-1-Flash](http://www.rundown.ai/tools/mai-cyber-1-flash?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=moonshot-lets-history-s-largest-open-model-loose) - Microsoft's cybersecurity AI for securing codebases 🐝 [Buzz](https://www.rundown.ai/tools/buzz?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=moonshot-lets-history-s-largest-open-model-loose) - Open-source workspace where AI agents join team chats 🎥 [FLUX 3](http://www.rundown.ai/tools/flux-3?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=moonshot-lets-history-s-largest-open-model-loose) **-** BFL’s multimodal AI with 20-second generations **Ilya Sutskever’s Safe Superintelligence** [announced](https://x.com/ssi/status/2081732119194394763?s=20&utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=moonshot-lets-history-s-largest-open-model-loose) a new partnership with Nvidia, including a “substantial investment” of reportedly $5B, to scale its research. **A Reddit user** [uncovered](https://www.reddit.com/r/ClaudeAI/comments/1v6fiyj/you_can_view_a_lot_of_shared_conversations_via/?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=moonshot-lets-history-s-largest-open-model-loose) a trove of Claude’s chats showing up in search engines, with Anthropic [attributing](https://techcrunch.com/2026/07/27/psa-your-claude-shared-chats-and-artifacts-may-have-ended-up-on-google/?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=moonshot-lets-history-s-largest-open-model-loose) the issue to users sharing the links and making them public. **Nvidia** is reportedly [discussing](https://www.wsj.com/tech/ai/nvidia-in-talks-with-openai-to-guarantee-250-billion-financing-for-data-center-3dd6eae3?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=moonshot-lets-history-s-largest-open-model-loose) a $250B financing guarantee to OpenAI to help facilitate the lease of a 10GW Ohio datacenter campus. **OpenAI CEO Sam Altman** [said](https://open.spotify.com/episode/1j1Mq3ouVslmPnWFCYPBJV?si=74d69f67cd1a4a5d&utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=moonshot-lets-history-s-largest-open-model-loose), “We are now in the singularity,” on the ‘Relentless’ podcast, saying this moment is the far-off dream he envisioned a decade ago. **Cursor** [launched](https://techcrunch.com/2026/07/27/cursor-makes-its-biggest-india-push-yet-ahead-of-spacex-acquisition-with-localized-pricing/?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=moonshot-lets-history-s-largest-open-model-loose) a new $7/month ‘Start’ plan in India to grow its footprint, providing only its own models to help developers build and ship with always-on cloud agents. **COMMUNITY** Every newsletter, we showcase how a reader is using AI to work smarter, save time, or make life easier. Today’s workflow comes from reader **Lyndal P.** in Brisbane, Australia: "After losing almost 30 kg, I realised none of the wardrobe advice online really fit my situation. I didn’t want to replace everything at once, and I certainly didn’t want to waste money buying clothes I’d regret. I started using ChatGPT as my personal stylist. I’d snap photos while I was shopping, and it would tell me honestly whether a colour suited me, which size would work, and how each piece would fit with what I already owned. Together we worked out my colour palette, built a wardrobe around clothes I genuinely love, and avoided plenty of expensive mistakes. The biggest surprise wasn’t just the outfits. It was how much more confident I became. Shopping went from feeling overwhelming to actually being enjoyable because I knew everything I bought had a purpose. My wardrobe finally feels like me." How do you use AI? Tell us [here](https://app.therundown.ai/community?submit=open&utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=moonshot-lets-history-s-largest-open-model-loose). Read our last AI newsletter: [Anthropic’s Opus 5 surprise](https://www.therundown.ai/p/anthropic-opus-5-surprise?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=moonshot-lets-history-s-largest-open-model-loose) Read our last Tech newsletter: [DARPA’s light-based quantum computer bet](https://tech.therundown.ai/p/darpa-big-bet-on-a-light-based-quantum-computer?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=moonshot-lets-history-s-largest-open-model-loose) Read our last Robotics newsletter: [A $1.7B computer for the physical world](https://robotnews.therundown.ai/p/travis-kalanick-1-7b-computer-for-the-physical-world?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=moonshot-lets-history-s-largest-open-model-loose) Today’s AI tool guide: [Make precise image edits on real photos with AI](https://app.therundown.ai/guides/make-precise-image-edits-on-real-photos-with-ai?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=moonshot-lets-history-s-largest-open-model-loose) RSVP to next workshop on July 30: [ChatGPT Sites for beginners](https://app.therundown.ai/live/an-intro-to-chatgpt-sites-for-beginners?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=moonshot-lets-history-s-largest-open-model-loose) See you soon, *Rowan, Zach, Shubham, and Jennifer — the humans behind The Rundown*