On 27 July, Moonshot AI released the open weights for Kimi K3, a language model with 2.8 trillion parameters that ranked third on the Artificial Analysis AI leaderboard at release, making it the most capable open model in the world to date. Starting today, we're making Kimi K3 available for all customers, running entirely on Swedish infrastructure.
Available for evaluation
Kimi K3 is available now for all our customers to test via our API. While we and our users validate performance under real-world load, we're collecting feedback and improving stability. Once the model passes our quality gates, we'll announce it as generally available.
curl https://api.berget.ai/v1/chat/completions \
-H "Authorization: Bearer $BERGET_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "moonshotai/Kimi-K3",
"messages": [{"role": "user", "content": "Your prompt here"}]
}'What happens to your data
Berget AI operates with zero retention. We never store prompts or responses. It's a direct application of GDPR's principle of data minimisation. Data enters the compute node, generates a response, and is removed before it's ever written to disk. No training on your inputs, no logging for later analysis, no data residency settings to configure. The data simply doesn't exist after the response is sent.
What sets K3 apart from other top models
Most models with equivalent performance are closed services. You send data via an API and get a response back, but have no control over the terms – and the vendor can change them or cut off access.
Kimi K3 is built on open weights. That means no single vendor controls access to the model. On Berget AI you run it through our OpenAI-compatible API: change the base URL and API key, and continue running on Swedish infrastructure under European law. The technology is the same, but the data never leaves the country.
Why this matters now
The release comes at a time when competition for AI dominance has escalated sharply. The US has invested over 900 billion dollars in AI infrastructure over six years, aiming to establish a global standard where value shifts from labour to licence fees. China, meanwhile, has launched a fifteen-year plan that puts open AI at the centre and subsidises access to models to strengthen industrial competitiveness.
Both paths create dependencies. The American dependency is commercial: a handful of companies control access to the most capable technology and can change the terms or cut off access. In June, the Trump administration ordered that the most capable models at the time be blocked for non-Americans. And for companies running on American platforms, US laws such as the CLOUD Act and FISA 702 apply regardless of where the servers sit. The Chinese dependency is geopolitical: technology subsidised by a state to build long-term influence.
Open models offer a third alternative. When the best open systems match the closed ones, and the EU has already made open source a cornerstone of its sovereignty strategy, the technical conditions exist to build independent infrastructure. American hardware, world-class open intelligence, operation on Swedish soil under European law – the combination is already real.
Kimi K3 is live for evaluation in our model catalogue.