The 230-Million-Parameter Surprise
A small non-transformer model matching rivals three times its size signals that capable private AI is arriving on ordinary hardware — while shadow AI becomes the leading incident cause.
29 June 2026 — Two numbers from this week deserve to sit side by side, because together they describe the actual state of AI for small business.
The first is 230 million — the parameter count of Liquid AI's LFM 2.5, a non-transformer liquid state-space model that matches transformer networks roughly three times its size. This is not a headline-grabbing frontier release. It is something more useful: proof that genuinely competent reasoning can run on modest local hardware, without an API call leaving the building. For a European SME, the implication is direct. Classification, routing, document triage and first-line response can now happen on-premises, at low cost, with no data crossing a border.
The second number is less comfortable. Cobalt's AI and Pentesting Pulse Report 2026 identified "shadow AI" — employees quietly using unapproved public AI tools — as the leading cause of corporate security incidents. Your staff are already pasting client documents into consumer chat windows. The report also found that AI features embedded in commercial software carry vulnerabilities that developers patch significantly more slowly than conventional flaws.
The distance between those two facts is the whole strategic picture. Capable private AI has become affordable in the same quarter that ungoverned public AI became the top incident category.
Meanwhile the geopolitical layer kept thickening. The US administration expanded national-security vetting, requiring individual approval for customers seeking access to Mythos 5 and newer OpenAI releases. Anthropic accused Alibaba of orchestrating a distillation attack — 25,000 fake accounts and 28.8 million API interactions harvesting Claude's reasoning into Alibaba's own models. And Goldman Sachs' June tracker reported roughly 11,000 net US jobs erased per month by generative AI, with junior knowledge roles disproportionately affected; Stanford's AI Index confirmed a sharp decline in employment of developers aged 22 to 25.
What this means if you run an SME. If you have not yet asked your staff which AI tools they use, this week is the week. The answer will be longer than you expect, and it will include at least one tool nobody approved. Once you know, you have a real choice: block, or replace with something private that does the same job. The second option is now genuinely available at a price a small business can carry.
Fourth in a weekly series on the developments that matter to small and mid-sized organisations.