Moonshot AI launched Kimi K3 on 16 July 2026, an open-weight multimodal reasoning model with 2.8 trillion parameters. On independent benchmarks the model currently ranks fourth among all frontier models, just behind GPT-5.6 Sol and the Claude top models. At the same time, the model's privacy position raises serious questions for Dutch professionals.
What makes Kimi K3 stand out?
The model uses a mixture-of-experts (MoE) architecture with 896 experts, of which only 16 are active per token. Despite the enormous parameter count, that makes it more efficient than its size would suggest. Other notable specs:
- Context window: 1 million tokens, included as standard in the API price
- Multimodality: text, images and video
- API price: $3 per million input tokens, $15 per million output tokens (cached input: $0.30/M)
- Open weights: available via OpenRouter; full release of the model weights expected on 27 July
The API via platform.moonshot.ai is already accessible, although there are signs of limited capacity with frequent 429 errors.
The GDPR problem
This is where it becomes relevant for Dutch users. Moonshot AI was founded in Beijing; its consumer service runs via Singapore, but the privacy policy does not name a specific storage country. When you call the API via platform.moonshot.ai, the model processes data in China, which is GDPR-critical.
Specifically, there are two issues:
- No opt-out for model training. The current privacy statement says user content is processed for "training and optimising models, among other things." Unlike ChatGPT, Claude or Gemini, Kimi has so far offered no in-product opt-out.
- Unclear transfers. Data may be transferred to servers outside the EU without the policy making clear which mechanism (adequacy decision, SCCs) applies.
Our safety profile overview shows how other models handle these kinds of questions.
What does this mean for your practice?
If you want to test Kimi K3, and that is understandable as the model is technically impressive, there are a few practical considerations:
- API via OpenRouter creates somewhat more distance from Moonshot's own infrastructure, but responsibility for the processing of your data remains with you.
- Do not enter personal data until Moonshot publishes a clear GDPR set-up (adequacy decision or SCCs with an explicit EU representative).
- Open weights (expected 27 July): if you self-host on EU infrastructure, you avoid the data transfer problem entirely. This does, however, require GPU capacity.
For sectors handling sensitive data, such as healthcare, legal and financial, we recommend using Kimi K3 for now only for non-personal, experimental tasks. See also our sector overviews for legal professionals and healthcare professionals.
Context: the wider market
Kimi K3 is arriving in a market that is moving at breakneck speed. GPT-5.6 Sol ($5/M in, $30/M out) and Grok 4.5 have recently launched; Claude Opus 4.8 is the current standard for complex reasoning. With its API price ($3/$15), Kimi K3 sits competitively between the mid-range and premium segments.
On our model overview we will keep prices and benchmarks up to date once they settle.
Sources: Kimi K3 tech blog – Moonshot AI · OpenRouter – Kimi K3 pricing · Privacy analysis Kimi, innFactory · Kimi privacy policy