The per-tier split I ran had no default: Small 3.2 for light work, Large 2512 for hard tasks, and DeepSeek on the fallback path. When Mistral discontinued Small 3.2, every role still on it had to migrate whether or not a test called for it. The fallback path rested on one model family, through one company. This is the ground behind Sovereign & local AI.
TL;DR
One EU default primary, Mistral Small 4, tuned per task through a binary reasoning toggle, replaces a model per tier; escalation stays cost-aware, and DeepSeek is dropped project-wide for a fallback family chosen for EU-provider redundancy rather than for the brand.
The decision, and the criteria I set first
The decision was how to allocate model choice across every task role and the fallback path. Four criteria came first.
- One model to operate. A per-tier split is costly to maintain, and a discontinuation forces the migration anyway.
- Per-task tuning without a model per tier. What varies is the task, not the layer.
- A hard cost bound. A stronger model is reached only on proven need.
- EU-provider redundancy on the fallback. It must not rest on one provider.
The cost bound bites at once: Mistral Medium 3.5 at $1.5/$7.5 per 1M tokens stays ruled out for tasks that do not need agentic-coding-tier capability.
Two ways to allocate model choice
Two allocations were live.
- A model per task tier: Small 3.2, Large 2512, and DeepSeek, no default primary.
- One default primary, tuned per task: Mistral Small 4 (
mistral-small-2603) across all four pipelines, its binary reasoning toggle set per task instead of a model per tier.
An earlier plan to make Large 2512 the default, and a model per tier, were both rejected. The digest role had moved off a Mandarin-trained dense model onto Mistral Small before this migration.
The fallback family: DeepSeek against Qwen
DeepSeek was the fallback incumbent, and it failed the redundancy criterion. Every EU router I researched, Cortecs and EUrouter, sourced DeepSeek from one company, Tensorix in Ireland. One company behind every routing option is one point of failure. The China-sovereignty risk compounds that finding; it does not lead it.
Qwen, Alibaba’s model family, was the alternative: qwen/qwen3.5-397b-a17b for heavy tasks, qwen/qwen3.6-35b-a3b for light, both genuinely redundant across Scaleway and OVHcloud.
A second, non-model fallback sits under it: the gateway keeps OpenRouter wired in as a permanent safety net, a router-level decision that belongs to the cutoff-survival test ranking every dependency.
The comparison
The allocations separate on maintenance and tuning, and agree on what a discontinuation forces.
| The criterion | A model per tier | One default primary |
|---|---|---|
| Models to operate | one per tier | one |
| Tuning per task | a different model | the reasoning toggle |
| Cost bound | by moving up a tier | only on proven need |
| A discontinuation forces | the migration anyway | the migration anyway, for one model |
The fallback pair separates on one criterion.
| EU-provider redundancy | DeepSeek | Qwen |
|---|---|---|
| Providers behind the researched routers | one (Tensorix, Ireland) | two (Scaleway, OVHcloud) |
What I chose, and why
Mistral Small 4 (mistral-small-2603) is the single default primary for every task role; Large 2512 is a candidate, never a role. It runs across all four pipelines (ai-newsletter-digest, lemonde-digest, dog-walk-scheduler, vps-monitor-digest), tuned per task through its reasoning toggle. The toggle is binary, not a fine dial: Mistral documents only "high" and "none", unsupported values may be silently coerced by the router, and the request carries a top-level "reasoning": {"effort": "high"|"none"} field. Large 2512 escalates only where Small 4 at "high" proves insufficient. DeepSeek is dropped project-wide for Qwen on the redundancy finding, with the China-sovereignty risk compounding it.
Across all 11 roles tested, at both efforts and all four pipelines, Small 4 proved sufficient every time; Large 2512 was never promoted.
What I gave up
- No per-tier specialization: one default trades a tuned model per layer for one model to operate.
- The agentic-coding tier stays out of the default path by cost.
- DeepSeek’s capability retires rather than sit on a one-provider path.
When the other option is right
- Where Small 4 at
"high"proves insufficient, Mistral Large 2512 is the per-task escalation. - Where a task genuinely needs agentic-coding-tier capability and the cost is justified, Medium 3.5 is the tool.
- Where a fallback path does not need two-provider EU redundancy, the choice narrows.
Each neighboring register is its own decision: which profiles run local, moving mail, docs, and calendar out of the cloud, on-device dictation, and routing tasks by shape rather than model.
Discuss the decision in your own context
The criteria decide the model; the brand does not. They travel to your pipelines: one default to operate, tuning without a new model, a cost bound, and a fallback no single provider can take down. If this is the shape of your problem, let’s talk.