Kimi K3 Weights Officially Released — A 2.8T-Parameter Open Model Closes In on the Frontier Camp
On July 27, Moonshot AI officially released the model weights for Kimi K3 under the "Kimi K3 License." It is a Mixture-of-Experts (MoE) model with 2.8T total parameters and 104B active parameters, making it the world's first open 3T-class model.
The benchmark results are what draw the eye: 93.5 on GPQA Diamond, 91.2 on BrowseComp, and 88.3 on Terminal-Bench 2.1, putting it at a level where it competes head-on with Claude Fable 5 and GPT-5.6 Sol. In particular, it shows an edge on long-horizon coding tasks with SWE-Marathon (42.0) and Kimi Code Bench 2.0 (72.9), and ranks top-tier in tool-integration capability with MCPMark-Verified (94.5).
On the architecture side, it adopts two new designs: Kimi Delta Attention (KDA) and Attention Residuals (AttnRes). With Stable LatentMoE, which selects 16 experts out of 896, the company claims roughly 2.5x better scaling efficiency compared to Kimi K2. Quantization-aware training for MXFP4 weights/MXFP8 activations is also built in from the SFT stage, with the expectation of running on a wide range of hardware.
It natively processes text, images, and video, and has a 1-million-token context window. The API is available at platform.kimi.ai, offering OpenAI/Anthropic-compatible interfaces.
Ilya Sutskever's SSI Announces Long-Term Nvidia Partnership — After Two Years in Stealth
Safe Superintelligence (SSI), founded by OpenAI co-founder Ilya Sutskever, has announced a long-term partnership with Nvidia. According to TechCrunch, SSI is aiming to scale up for its next phase after roughly two years in stealth.
SSI is a startup on a mission to build "safe superintelligence," a new venture Sutskever launched drawing on his experience at OpenAI. The detailed terms of the Nvidia partnership have not been disclosed, but securing compute infrastructure appears to be the primary aim. Given how critical GPU access is in AI research, a direct partnership with Nvidia carries strategic significance for frontier development.
Cognizant Becomes an Anthropic Global Premier Partner — Enterprise Claude Deployment Gets Serious
Anthropic has expanded its partnership with the IT services giant Cognizant, announcing that Cognizant has become a Global Premier Partner in the Claude Partner Network.
Cognizant already has more than 30,000 employees who have completed Claude training, and it has embedded Claude into its own platforms, including Flowsource™ (software engineering), Neuro® AI Engineering, and Neuro® IT Ops. In Flowsource's Spec-Driven Development module, Claude Code operates based on specifications, coding standards, and architecture design, with outputs evaluated before reaching production.
Client results cited include a manufacturing customer experience portal (built in six months), a biopharma contract review system (40% reduction in review time, over 88% extraction accuracy), and a risk navigation tool for the insurance industry (eight hours saved per week). Cognizant CEO Ravi Kumar S's observation that the bottleneck to AI adoption lies in "industry context and engineering scale" neatly captures the reality of enterprise AI adoption.
Microsoft Research Releases Fara 1.5-27B — Operating a Browser from Screenshots Alone
Microsoft Research AI Frontiers has released the browser automation agent Fara 1.5-27B on Hugging Face.
Fara 1.5-27B is a Computer Use Agent (CUA) built on Qwen3.5-27B, notable for operating on web browser screenshots alone as its input. It does not rely on the DOM or the accessibility tree, making judgments from the same visual information a human user sees. Because it directly predicts pixel coordinates to specify click and drag targets, no separate grounding model is needed.
It is designed to complete multi-step tasks end to end, such as form filling, bookings, shopping cart checkout, and trip planning, and it is fine-tuned on data generated by a multi-agent pipeline called FaraGen1.5. The recommended deployment is in combination with MagenticLite.
Google AI Search Reaches 43% of All Searches — AI Overviews Become the "De Facto" Standard
According to new data reported by TechCrunch, Google AI Overviews now appear on 43% of all searches — a figure showing that AI-generated answers are becoming the default way people access information.
This shift represents a structural change in search behavior. The transition is accelerating away from a model where users find answers themselves from a list of links, toward one where they directly receive AI-summarized answers. For content creators and the SEO industry, it signals that a fundamental change in traffic sources is imminent.
Nvidia's $750B Deals Stir Renewed Concerns Over "Circular Financing"
Bloomberg's report on Nvidia's roughly $750 billion in deals has reignited concerns within the AI industry over "circular financing."
The structure — AI companies buying Nvidia GPUs, with that money flowing back to them through Nvidia's AI investments — could, depending on your perspective, become a self-amplifying mechanism for the AI bubble. A cycle in which Nvidia's revenue is propped up by investments in AI companies, and those AI companies in turn buy GPUs from Nvidia, raises questions about sustainability. That said, the details of the actual deals and the flows of money are complex, and a definitive verdict is best avoided at this point.
Takeaways
What today's news reveals is a trend: as AI model performance converges, differentiation is shifting toward "openness," "enterprise integration," and "ecosystem building." Kimi K3 offers frontier-class performance with open weights, while Microsoft demonstrates practical utility in a concrete use case — browser automation. At the same time, Google AI Search's adoption rate and the scale of Nvidia's deals throw the fundamental question of the economic sustainability of AI infrastructure into sharp relief once again.