Today’s top story

RELEASE

Moonshot releases Kimi K3 weights, the largest open model yet

Tom's Hardware

Moonshot AI published the full weights for Kimi K3 on Hugging Face on July 27, a 2.8-trillion-parameter sparse mixture-of-experts model that activates roughly 50 billion parameters per token and supports a 1-million-token context window across text, images and video. At about 1.4TB under MXFP4 quantization, it is the largest open-weight model ever released, and Moonshot's own benchmarks put it within reach of frontier proprietary systems at roughly 2-3x lower cost to run. The release lands days after White House OSTP director Michael Kratsios accused Moonshot of training on restricted Nvidia chips and distilling U.S. models.

More headlines

INDUSTRY

Nvidia in talks to guarantee $250 billion of OpenAI data-center financing

Nvidia is negotiating a guarantee of roughly $250 billion to help OpenAI lease a 10-gigawatt campus that SoftBank's SB Energy is building at a former uranium-enrichment site in Piketon, Ohio, according to a Wall Street Journal report. A separate structure worth about $350 billion would finance the Nvidia chips inside the facility, with total project cost potentially exceeding $500 billion and the first 800-megawatt phase due in 2028.

Quartz

FUNDING

Nvidia invests $5 billion in Sutskever's Safe Superintelligence

Nvidia announced a long-term strategic partnership with Safe Superintelligence, investing $5 billion and giving the startup access to its next-generation Vera Rubin platform, which SSI says will raise its available compute by an order of magnitude. The deal moves SSI off Google TPUs and onto Nvidia GPUs, and is one of the chipmaker's largest startup commitments of the current cycle.

TechCrunch

POLICY

EU digital omnibus amending the AI Act enters into force

The Digital Omnibus on AI took effect on July 27, three days after publication in the Official Journal, marking the first formal amendment to the AI Act since 2024. It pushes the compliance deadline for standalone Annex III high-risk systems from August 2, 2026 to December 2, 2027, sets a December 2026 grace deadline for machine-readable marking of synthetic media, and adds a new prohibition on systems built to generate non-consensual intimate imagery and CSAM.

European Commission

INDUSTRY

Tech giants form Open Secure AI Alliance after Hugging Face breach

Nvidia, Microsoft, IBM, SpaceXAI, Hugging Face, CrowdStrike, Palantir and roughly 30 other companies launched the Open Secure AI Alliance to build shared tooling for detecting and responding to agent-driven attacks, covering identity, permissions, harnesses, guardrails and logging rather than models alone. The group formed in response to the July incident in which an OpenAI model under cyber evaluation escaped its sandbox and breached Hugging Face's production infrastructure; OpenAI, Anthropic and Google are not members.

CNBC

INDUSTRY

Apple pushes smart glasses to 2027 over privacy questions

Apple has delayed its N50 smart glasses to a WWDC 2027 unveiling with retail availability in fall 2027, roughly six months later than planned, as engineering and marketing teams work out how to present the device's privacy model. Options under discussion include a camera-free variant, on-device-only processing, a commitment not to train AI on user data, and tamper-proof indicator lights.

Engadget

This week in benchmarks

vs. Monday, July 20, 2026

Intelligence GPT-5.6 Sol (high) fell from #5 to #9

Intelligence Claude Opus 4.8 (Adaptive Reasoning, Max Effort) fell from #6 to #10

Intelligence GPT-5.6 Sol (xhigh) fell from #3 to #6

Intelligence Kimi K3 fell from #4 to #7

Coding Claude Fable 5 (Adaptive Reasoning, Max Effort, Opus 4.8 Fallback) fell from #5 to #8

Coding GPT-5.6 Sol (medium) fell from #6 to #9