AI Radar: xAI Grok 4.3 — Fast, Closed, Frontier
xAI is moving fast. With Grok 4.3 and the multi-agent offshoot 4.20, the lab shows where proprietary frontier models are heading — which is exactly why it belongs on the radar, even though it does not fit our self-hosting approach.
The State of Play
- Grok 4.3: available since April, with native video input, document generation (PDF, spreadsheets, slides), and noticeably better tool-calling.
- Grok 4.20: a multi-agent architecture with four cooperating agents and a 2M-token context, plus real-time access to X data.
- Grok 5: training on the Colossus 2 cluster, in the multi-trillion-parameter range — expected later this year.
- Image and video: with grok-imagine, generative video models are in the API too.
The Caveat
Grok is consistently closed: no open weights, no self-hosting, data runs through xAI and is tightly interwoven with the X platform. For confidential business data — and for anyone who takes data sovereignty seriously — that is a disqualifier, no matter how fast the model is.
Our Take
We track Grok as an indicator of pace and features: what xAI demonstrates often shows up months later as an open alternative elsewhere. It still does not belong on our self-hosting path. Speed is no substitute for data sovereignty — and in the SME world, sovereignty is the yardstick.