Stock these open models before political disruption hits

What the two reports say

Two political-risk reports, written three years apart, point the same direction. Kroll’s enterprise-security team warned in January 2023 that post-election instability in democratic countries was becoming routine; the Tony Blair Institute’s six-country voter survey published the receipts in January 2025. Traditional centre-left and centre-right parties have collapsed from about three-quarters of the vote at the start of the century to just over half today, with volatility now a permanent feature of our politics.

From three-quarters of the vote to just over half today — the share going to traditional social-democratic and centre-right parties across Western democracies, per TBI analysis of Deltapoll polling of 12,000 voters across six countries.

Kroll calls the emerging pattern a “playbook”: pre-election fraud claims, refusal to accept results if the candidate loses, mass protests escalating into attacks on government infrastructure. They note that the period in which elections can result in political instability in democratic societies is expanding, and that social media makes the playbook easier to repeat.

What this means for cloud AI

None of this is directly about AI. But cloud AI sits behind three choke points at once: a handful of US hyperscalers, a small group of frontier-model labs, and the export-control machinery that decides who can use what. The last twelve months have shown how quickly access can move. The US ordered Anthropic to pull Fable 5; GPT-5.6 Sol shipped under restricted US access. Neither move was about your business; both changed what you could buy on Monday.

The hedge is old-fashioned: own the thing. Open-weight models — the actual neural-network files — downloaded to a machine you control and run through a local model runner do not need anyone’s permission, sanctions licence or API key. If a provider cuts you off, your shelf still works.

The shelf, by job

For a single high-end workstation or a second-hand dual-GPU rig, you can comfortably hold half a dozen serious models. The right mix depends on what you do, but a sensible default:

• Coding and agentic work — Qwen3.6-35B-A3B is the established local coding agent at this size; it pairs cleanly with Claude Code as a backend. GLM-5.2 is the alternative if you want a model that has tied Opus on real coding tests.
• General chat, drafting and reasoning — MiniMax M3 is the open-weight frontier small teams can actually run, with usable speeds on consumer Apple hardware. Mistral Medium 128B covers heavyweight reasoning jobs.
• Creative and longer-form writing — DeepSeek V4 Flash is the strongest open-weight writer at the mid-tier; Step 3.7 Flash is the cheaper substitute when speed matters more than peak quality.
• Cheap utility, classifiers, routing — Gemma 4 E2B runs three jobs from a tiny footprint and earns its keep on triage, intent detection and moderation.

How to stock the shelf this weekend

You don’t need every model above. One per row covers most of what a small firm will throw at a local stack. A weekend is plenty:

• Saturday morning (90 minutes). Pick a runner — Ollama if you’re CLI-comfortable, LM Studio if you’d rather click. Install it. Pull two models: Qwen3.6-35B-A3B at Q4 and Gemma 4 E2B at Q8. That’s your coding-and-utility pair, and it’s all most teams need for the first month.
• Saturday afternoon (90 minutes). Wire one into a real workflow. Easiest wins: a local classifier replacing a paid intent-detection API, or a local coding agent pointed at a non-production repo. The Gemma 4 fine-tune for marketing copy is a worked example of what a small utility does in practice.
• Sunday (60 minutes). Add the heavyweight. Pull MiniMax M3 or Mistral Medium 128B at the quantisation that fits your RAM. Run the same prompts you’ve been sending to a paid API and write down what changed. Cost-per-task should drop to a rounding error; latency will probably get worse — that’s the trade.

You’re not building a bunker. You’re building a shelf. If the political weather turns, your shelf still has weights on it.

What to watch

Three signals worth tracking over the next six months:

• A US executive order, sanctions notice or export-control update that names a model family you depend on — the Fable 5 pull and GPT-5.6 Sol restricted-access episodes are the template.
• A UK regulatory move that pulls data-residency lines tighter; sovereign-model projects like Lumen Sovereign and the £500M Sovereign AI Unit become more attractive the moment compliance tightens.
• A multi-day outage at a single hyperscaler. Local weights keep the lights on; cloud-only stacks don’t.