The way large models are sold to regulated industries has just changed, and most buyers will miss it. The headline number on every release is the total parameter count. The number that decides whether a public administration, an aerospace integrator, or an industrial operator can self-host a frontier-class model on its own hardware is the active count per pass.
Active parameters set the inference bill, and therefore the deployment decision.
Aleph Alpha's Kolibri is the cleanest recent case: 78 billion total, 3 billion active per pass, English and German, 1M tokens of context, and an Apache 2.0 license on Hugging Face that lets a buyer download every weight. The 78B fits the industry pitch. The 3B is what decides whether a regulated team can run the model inside its own jurisdiction without renting from a US frontier API.
Read every future MoE release this way. The total fits a category that already includes Llama 3.1, Mistral, Qwen, and Gemma. The active count, the license, and the language coverage are what a regulated buyer actually evaluates. "Sovereign" is Aleph Alpha's framing for the supply chain, not a third-party certification; treat it as a testable claim, not a label.
Reported by Sky for Type0, from Kolibri Has Landed: A Sovereign Open-Weight Model. Read the original: aleph-alpha.com