Nvidia forms the Open Secure AI Alliance + Open-weights letter
On Monday July 27, Nvidia announced the Open Secure AI Alliance with founding members Adobe, CrowdStrike, Hugging Face and Dell Technologies to develop and share tools for AI safety and cybersecurity.
The framing: Nvidia argued “Blanket restrictions on open frontier AI systems would weaken defensive capacity and risk concentrating power, dependence and vulnerability in a few closed providers”.
Just before that, on July 24, Nvidia was among signatories of “Open Weights and American AI Leadership,” urging policymakers to support open-weight AI models — defined as systems with publicly accessible core components and open-source tools enabling inspection and collaborative improvement. Nvidia contributed its new open-source Object-Oriented Agent project on GitHub to help test, track, review and regulate agent actions.
Anthropic notably did NOT sign the letter, then published its own position on open-weights on July 27.
Open-weight shockwave — DeepSeek V4 and Kimi K3
The biggest market mover: China’s Moonshot AI.
Kimi K3: Launched July 16-17 as a 2.8T-parameter MoE model (fires 16 of 896 experts per token), with 1M-token context, native vision, and #1 on Arena’s Frontend Code leaderboard. It is the largest open-weight model to date, surpassing DeepSeek V4 Pro’s 1.6T and Llama 4 Behemoth’s 2T. Weights are scheduled for public download July 27 at 1,561 GB.
Why Wall Street panicked: The open-source model is “one of the most efficient AI models ever released, and its performance is said to be near or on par with top U.S AI models”. Nasdaq 100 had its worst July in over two decades on the news, echoing the January 2025 DeepSeek R1 selloff that wiped $1T in a day.
Pricing context: Kimi K3 API is $3 input / $15 output per million tokens — far below Fable 5’s $50 output, but above DeepSeek V4 Pro’s $0.18 blended.
DeepSeek V4 Pro also solidified its position as the efficiency king, scoring 80.6% on SWE-bench Verified at release.
Moonshot AI releases Kimi K3 (2.8T parameters)
Now the top performer on major coding benchmarks, surpassing Western labs in several categories.
2.8T parameters places it among the largest publicly disclosed models.
Signals China’s increasing competitiveness in frontier-model capability.
Thinking Machines ships new open-weight flagship model
Strengthens the open-weight ecosystem, giving enterprises more options outside proprietary labs.
Focus on reproducibility, transparent training data, and controllable inference behavior.
