Moonshot AI released Kimi K3 on July 16, marketing it as the first open 3T-class model: a sparse mixture-of-experts system with 2.8 trillion total parameters that activates 16 of 896 experts per token, a 1-million-token context window and native vision.

Two architecture changes carry the claims — Kimi Delta Attention for scaling across sequence length and Attention Residuals for selective depth-wise retrieval — which Moonshot says yield roughly 2.5x better scaling efficiency than Kimi K2, a vendor figure that has not been independently reproduced. It ships always-on max reasoning, prices at $0.30 per million cache-hit input tokens, $3.00 cache-miss input and $15.00 output, and is live on Kimi.com, Kimi Work, Kimi Code and the API. The catch: the weights, promised under a new Kimi K3 license, were not downloadable until July 27, so the largest-open-model-ever billing was unverifiable at launch. Moonshot's benchmark footnotes also lay out exactly which agent harness — Kimi Code, Claude Code or Codex — scored what.

The European Commission issued two sets of binding specification measures to Google under the Digital Markets Act on July 16. The first requires that third-party AI assistants get the same Android access as Gemini: users must be able to wake their chosen assistant by voice, similar to Hey Google, and delegate actions inside apps — booking a taxi, suggesting chat replies, answering questions about a recently visited place.

The second specifies how Google shares Search data with rival search engines, including making AI chatbots with search features eligible, with a multi-layered anonymization method and a published formula for pricing access. Google must start sharing search data from January 2027, and the Android changes reach users from July 2027. Google's global affairs chief Kent Walker called the measures unprecedented risks to user privacy, device security and national security. Because these are specification proceedings rather than non-compliance investigations, they carry no fines — but they are legally binding.

Mira Murati's Thinking Machines Lab published Inkling on July 15, its first foundation model: a mixture-of-experts transformer with 975 billion total parameters and 41 billion active, up to a 1M-token context, pretrained on 45 trillion tokens of text, images, audio and video.

It was trained from scratch with encoder-free audio and vision inputs, shipped under an Apache 2.0 license, and paired with the company's Tinker fine-tuning platform at 64K and 256K context options. The framing is unusually blunt: the blog states Inkling is not the strongest overall model available today, open or closed, and positions it as a customizable base rather than a leaderboard winner. On agentic coding the company claims Inkling hits the same Terminal Bench 2.1 performance as Nvidia's Nemotron 3 Ultra using one third the tokens, and the model exposes a controllable thinking effort dial. WIRED notes Thinking Machines used Inkling to fine-tune and improve itself, and that its chain of thought got terser over training while outputs stayed the same.

TechCrunch reported on July 13, citing three sources, that Nous Research is finalizing a round of at least $75 million at a $1.5 billion valuation, led by Robot Ventures with significant participation from Union Square Ventures.

The company had raised roughly $70 million cumulatively before this, including a $50M Series A led by Paradigm, so the new round alone is larger than everything that came before it. The pitch hinges on Hermes, the open-source agent framework: the company says it can learn from usage and build skills without manual intervention, and users run it on a desktop or VPS, chatting with agents through Telegram and Discord. GitHub numbers cited in the report put Hermes at roughly 214,000 stars and nearly 40,000 forks.