Kimi K3 Arrives: The World's Most Powerful Open-Weight Model Goes Public
Moonshot AI's 2.8 trillion-parameter model supports 1 million tokens of context and native vision, with weights released under an MIT-inspired license. Every major inference platform announced Day-0 support within hours.
Moonshot AI released Kimi K3, the strongest open-weight model ever made available to the public. Built on a Mixture-of-Experts architecture with 104 billion active parameters out of 2.8 trillion total, the model supports a 1-million-token context window and native visual understanding. It achieves near-frontier benchmark performance while being 2 to 3 times easier to run than comparable proprietary models, thanks to innovations like Kimi Delta Attention and ultra-sparse MoE routing. On Hugging Face, Kimi K3 received over 4,000 likes within 30 minutes of release — the fastest adoption rate in the platform's history. Within hours, vLLM, SGLang, Together AI, Fireworks, Modal, Baseten, Nebius, and DigitalOcean all confirmed Day-0 inference support. The model is also integrated into Cursor, Ollama Cloud, and Vercel AI Gateway with zero data retention agreements. Moonshot AI simultaneously open-sourced several companion projects: MoonEP, a high-performance communication library for distributed MoE workloads; AgentENV, a distributed system for running agent environments at scale; and FlashKDA, a CUTLASS-based attention kernel achieving 1.72x to 2.22x prefill speedup on H20 GPUs. The Kimi K3 license, inspired by MIT, restricts commercial use for companies exceeding $20 million in annual revenue, requiring a separate commercial agreement and attribution above 100 million users. teortaxesTex estimated the training run exceeds 2e25 FLOPs, likely the largest model ever trained in China.
Anthropic Clarifies: We Never Sought to Ban Open-Weight Models
CEO Dario Amodei stakes out a nuanced position — open weights are a public good, but safety testing must apply to all sufficiently powerful models, open and closed alike.
Anthropic CEO Dario Amodei published a detailed statement clarifying the company's position on open-weight models, explicitly stating that Anthropic has never advocated for banning them. He characterized open-weight models without dangerous capabilities as genuine public goods that advance scientific progress. However, Amodei rejected protectionist bans as ineffective substitutes for targeted national security measures. He proposed three concrete policy interventions: restricting powerful chips from reaching authoritarian regimes, stopping industrial-scale model distillation by competing powers, and mandating rigorous pre-deployment safety testing for all models — both open and closed — that cross capability thresholds. In a notable divergence from the open-weight joint letter signed by NVIDIA, Microsoft, and over fifty others, Amodei challenged the assertion that openness is inherently safer, arguing that an attacker-defender asymmetry could make open models more useful to malicious actors. The answer, he wrote, is empirical evaluation, not ideological positioning. Nathan Lambert responded that the article rehashed existing positions without adding anything new, and called the distillation ban proposal misguided. teortaxesTex was more blunt, suggesting Anthropic has no profound advantage in training science — only vast compute capacity and a head start.
Ilya Sutskever Signals SSI Scaling With Major NVIDIA Investment
The secretive AI lab has completed internal verification and is preparing for a tenfold increase in computing power over the next twelve months.
Ilya Sutskever broke his characteristic silence with a brief but consequential message on social media: "Time to scale that SSI." The statement followed reports that NVIDIA has made a major new investment in the company, which will increase its compute capacity by 10x over the next 12 months. Internal verification has been completed, and SSI is now preparing to scale its training infrastructure to unprecedented levels. Notably, the company has yet to release any public model, making it one of the most closely watched but opaque organizations in the AI landscape. Among the new wave of AI labs, Sutskever's SSI is widely considered the most promising — combining vast resources with one of the field's most celebrated research minds. The compute expansion suggests that whatever SSI is building internally, its training runs are about to become materially larger than anything the lab has attempted before, raising questions about what the former OpenAI chief scientist has been building behind closed doors.
Let's stop believing the PR that closed models are safer. That's just regulatory capture.
Andrew Ng, responding to NVIDIA's open-weight joint letter
Hugging Face and NVIDIA Form Open Secure AI Alliance
Hugging Face joined NVIDIA and other industry leaders to launch the Open Secure AI Alliance, an initiative focused on identifying and addressing software vulnerabilities in AI systems through open-source collaboration. Adobe also announced its participation, signaling broad industry support for shared security infrastructure. The alliance reflects growing recognition that AI security is a collective challenge rather than a competitive advantage.
Grok 4.5 Becomes Most Cost-Effective Cybersecurity AI Model
In the deepsec benchmark for full-repository security review, Grok 4.5 emerged as the leading model on price-performance. It is 10x cheaper than Sol, 5.7x cheaper than Opus 5, and 2.2x cheaper than Kimi K3 while delivering performance competitive with K3. Sol retains the absolute frontier position, while Opus 5 follows. The benchmark, which runs on user-owned infrastructure with sandboxed execution, tests models on finding real-world vulnerabilities across entire codebases.
Kimi K3 Arrives in Cursor, Scores Near Frontier on CursorBench
Cursor integrated Kimi K3 via US-based inference from Fireworks, Together, and Baseten, with zero data retention guarantees. The model scores near the frontier on CursorBench, giving developers access to state-of-the-art capabilities directly in the editor. Vercel AI Gateway also launched K3 access with similar zero data retention agreements signed with all providers.
Kaifu Lee: Companies That Didn't Sign the Open-Weight Letter Matter More
Kaifu Lee pointed out that the list of AI companies refusing to sign the open-weight joint letter is more revealing than the signatory list. He specifically noted Claude AI's absence, framing the divide between open and closed model camps as increasingly defining for the industry's future direction. The letter, led by NVIDIA and Microsoft, attracted nearly every major Western AI company, with one prominent holdout drawing widespread attention.
OpenAI Expands GPT-Live Voice to Enterprise Users Globally
GPT-Live, OpenAI's real-time voice interaction feature for ChatGPT, is now available for Education, Business, and Enterprise plan subscribers worldwide. The expansion marks a significant step in OpenAI's push into enterprise communication, as live voice becomes an increasingly competitive feature category across all major AI platforms.
Cohere Launches North Automations for Plain-Language Workflows
Cohere released North Automations, allowing users across all technical levels to design automated workflows using natural language. The feature enables non-technical employees to build AI-powered processes without writing code, reflecting a growing industry trend toward making AI accessible beyond engineering teams.
Moonshot AI Open-Sources MoonEP Communication Library
MoonEP is a high-performance communication library that reduces overhead in expert-parallel communication for large-scale MoE training and inference. By optimizing how experts communicate across distributed systems, MoonEP directly addresses one of the main bottlenecks in scaling sparse models like Kimi K3.
AgentENV Powers the RL Training Behind Kimi K3
In collaboration with kvcache-ai, Moonshot AI open-sourced AgentENV, a distributed system for running agent environments at scale. Supporting fast snapshot, resume, and fork of parallel workflows, AgentENV was the infrastructure backbone for Kimi K3's agentic reinforcement learning training, showing how core model capabilities increasingly depend on robust agent simulation environments.
FlashKDA Delivers 1.72x–2.22x Faster Attention on H20 GPUs
FlashKDA, a CUTLASS-based implementation of Kimi Delta Attention kernels, achieves 1.72x to 2.22x prefill speedup over the flash-linear-attention baseline on NVIDIA H20 GPUs. It works as a drop-in backend replacement for flash-linear-attention, making it immediately usable in production inference stacks.
Ex-Anthropic Employee: Hackers Prefer Subsidized APIs Over Open Models
Amjad Masad, CEO of Replit, cited a former Anthropic employee's observation that malicious actors overwhelmingly use heavily subsidized lab APIs for cyberattacks rather than open-weight models, directly challenging a core argument made by closed-model proponents — that open weights uniquely enable misuse.
NVIDIA's AI Translation Platform Cuts Processing Time by 70%
NVIDIA's digital marketing team built an internal AI translation platform using Nemotron Speech and the NeMo framework, trained on seven years of marketing data. The platform reduced translation processing time by approximately 70% per request and lowered annual costs by about 25% through consolidation of legacy integration plugins and vendor optimization.
teortaxesTex Proposes Air-Gapped Evaluation for AI Safety
In response to concerns about AI models escaping sandboxes, teortaxesTex proposed a straightforward solution: air-gapped evaluation clusters. He argued that major labs like OpenAI have the resources to implement this and should stop treating the problem as intractable when a physical air gap is an established security practice.
Kimi K3 Breaks HuggingFace Growth Records
HuggingFace CEO Clement Delangue reported K3 received over 4,000 likes within 30 minutes of launch, calling it the fastest release growth in the platform's history.
Nathan Lambert: Anthropic Open-Weight Article Adds Nothing New
Lambert called Anthropic's position statement a reasonable but unsurprising repeat of existing stances, reiterating that banning distillation remains a misguided proposal.
Mistral CEO Arthur Mensch Joins Open-Weight Alliance
Mensch announced Mistral joined the open-weight alliance, arguing open models are essential for digital safety and US competitiveness.
Kimi K3 Available on Ollama Cloud With Claude Code Integration
Ollama launched K3 for Pro and Max subscribers, including support for Claude Code via the command ollama launch claude --model kimi-k3:cloud.
OpenAI Studies AI as Generalist Tool for Small Teams
New research shows AI helps small businesses work across functions, letting non-specialists solve problems outside their job descriptions.
Chollet: External Markets as Reward Signals for Agent Evaluation
François Chollet highlighted the concept of using real economic interactions as reward signals, fundamentally changing the agent optimization problem.
K3 Paper: Agents Can Crash Host Machines Through Kernel Panics
Rauch noted that Kimi's paper shows container-level isolation is insufficient for agent security — agents caused kernel panics, but Firecracker microVMs remained safe.
Nathan Lambert Breaks Down Kimi K3's Commercial Restrictions
The K3 license, inspired by MIT, requires companies with revenue over $20M annually to negotiate commercial terms and display attribution above 100M users.
Claude Design Called the Most Underrated AI Product
McKay Wrigley praised Claude Design with Opus 5 as transformative, saying it should be 100 times more popular.
AI-Written YouTube Script Gets 20x the Creator's Normal Views
VentureTwins found an AI-scripted video outperformed human-written content dramatically, with commenters enthusiastic and unaware.
K3 Tech Report Omits Training Tokens, Signals Massive Scale
teortaxesTex notes the K3 report doesn't disclose token count, but estimates compute well over 2e25 FLOPs — China's largest model to date.
Anthropic May Have a Larger Internal Teacher Model
teortaxesTex reports Anthropic likely has an internal teacher model larger than Mythos-Preview, with the currently served model near K3's scale.
Nathan Lambert Uses Fable Agent to Auto-Build Research Slides
Lambert demonstrated AI agents fetching data from wandb API to build presentation slides, showing automated research communication.
Midjourney V8.2 Introduces Highly Novel Visual Styles
Users praise Midjourney V8.2 for a wide range of unique, creative visual styles beyond what other image generation models offer.
Recraft V4.1 Turns Mockups Into Photorealistic Campaign Visuals
Recraft V4.1 generates natural lighting, believable materials, and subtle imperfections indistinguishable from real advertising shoots.
PixVerse Launches Avatar Lip-Sync From Still Images
The Avatar Lip Sync feature generates natural lip-synced speaking videos from user-uploaded images or short video clips.
Swyx: Claude Code Was Open-Sourced and Nothing Happened
Swyx observed that Claude Code's accidental open-sourcing earlier this year had approximately zero impact on competitors or the ecosystem.
Sakana AI Signs Open-Weight Joint Letter
Co-founder Hardmaru stated open ecosystems form the foundation of a healthy AI industry, joining the coalition of signatories.