Two releases this week deserve attention if you work professionally with AI: the largest open-source model yet, and a new way to let AI use login credentials safely.
Kimi K3: the largest open model to date
On 16 July 2026, Chinese company Moonshot AI launched Kimi K3, a model with 2.8 trillion parameters. For comparison, the previous open record was just under a trillion. K3 uses a Mixture-of-Experts architecture in which only 16 of the 896 experts are active per token, keeping inference costs manageable despite the scale.
What makes the model interesting for Dutch professionals:
- 1 million token context window, useful for analysing large document sets or long contracts.
- Native vision: processes images and charts without a separate call.
- Open weights are released on 27 July 2026, which makes on-premises deployment possible and is therefore more relevant to anyone wanting to get around the data location question.
API prices via Kimi are $0.30/MTok (cache-hit input), $3.00/MTok (cache-miss input) and $15.00/MTok output. That is more expensive than Claude Sonnet 5 or GPT-5.6 Luna, but cheaper than the most powerful closed models. Benchmarks place K3 at the level of the current frontier, although independent validation on European language averages is still lacking.
There is one serious caveat: Moonshot AI is a Chinese company. The API processes data on servers outside the EU. If you act as a controller under the GDPR, you need a solid legal basis for that transfer. The open weights offer a way out: once you host the model yourself, on an EU server, the jurisdictional risk disappears. You can read more about data location per model on the models page.
Source: Kimi K3 Tech Blog – Moonshot AI · VentureBeat – China's Moonshot AI releases Kimi K3
1Password for Claude: logging in without the AI seeing your password
On 17 July 2026, 1Password and Anthropic released a browser integration that allows Claude to carry out agentic browser tasks, including logging in to websites, without the passwords ever reaching the model's context.
How it works: when Claude lands on a login page, 1Password explicitly asks the user for approval on which credentials may be used. Once approved, 1Password fills in the credentials directly into the form, outside the model's field of view. When the task is finished, access expires automatically.
For organisations: on Team and Enterprise plans, Agentic Mode is switched off by default; only owners can activate it. Credit card details and identity documents from the vault are for now not available to the AI agent.
This is relevant for anyone deploying AI agents for repetitive web-based tasks, such as filling in portals, retrieving reports or managing account settings. The zero-exposure principle limits the risk of login details ending up in prompt logs, server memory or training data, a point the Dutch DPA (AP) also stresses in its recent guidance: personal data must not end up outside the defined processing purpose.
Available now for Mac, for individual, family and business plans; requires the 1Password desktop app, the browser extension, the Claude desktop app and the Claude browser extension for Chrome.
Source: Help Net Security – Claude can now sign into websites with 1Password · 1Password Blog – 1Password for Claude
What it means for you
Kimi K3 widens the range of choice for anyone wanting to deploy open models, especially if the weights are run in-house. The 1Password integration is a concrete step towards making agentic AI safer, a category that is growing fast but still needs to mature in terms of security standards.
Both developments are symptomatic of a broader shift: AI tools are not only becoming more powerful, but also gradually easier to fit into work processes with real privacy requirements.
