The three levels
CoeOS is the same at every level: the same client, the same box, the same routing by skill. What changes is where confidential work goes and what stays in the cloud. A level is not a fixed product: it is a set of settings.
At a glance
Section titled “At a glance”| Open | Hybrid | Local | |
|---|---|---|---|
| Confidential goes | your choice: no distinction, or EU providers under DPA + ZDR | to a model hosted on your premises | stays on site |
| The rest goes | to the cloud, best model per axis | to the cloud, best model per axis | stays on site |
| What switches | the Confidential switch | the Confidential switch | nothing: everything stays on site |
| Hardware on your side | none | one Apple Silicon server | an Apple Silicon cluster |
The best open models, in the cloud. Every request goes to the model that scores highest on its axis on the TMB benchmark.
You choose the guarantees:
- No privacy or retention constraint: CoeOS goes through OpenRouter. On the TMB scoreboard of 28 October 2026, CoeOS scores 98.4/100 at $0.039 per test, against 98.0 at $0.394 for Fable 5: 10% of the price for a higher score.
- With guarantees: European providers under a DPA (data processing agreement) and ZDR (zero data retention, nothing used for training). The price depends on the providers you choose.
Hybrid
Section titled “Hybrid”Confidential work goes to a model hosted on your premises, on an Apple Silicon server running OdyssAI-X, OdyssAI’s inference engine. The rest keeps the power of the cloud.
Every conversation has a Confidential switch, on by default. A conversation that sent a confidential turn never goes back to open. See confidentiality.
Everything stays on site, no compromise. Three models, three jobs, and CoeOS routes between them as it does in the cloud:
| Model | Job | Reference |
|---|---|---|
| GLM 5.3 | precision: planning, calculation, business writing | TMB 94.1/100, local Q8 |
| MiMo V2.6 Flash | speed: everyday work, summaries, tool calls | 500+ tok/s aggregate |
| Qwen3 Next | scale: fully local agentic flows, large teams | up to 1,000 tok/s aggregate |
Aggregate throughput measured in Replica mode on Mac Studio M3 Ultra, short answers (150 tokens max). Four 256 GB nodes: MiMo V2.6 Flash 518 tok/s at 96 concurrent requests, Qwen3 Next 8-bit 985 tok/s at 256. Three 96 GB nodes: Qwen3 Coder Next 6-bit 650 tok/s at 96. A single request runs slower; a team runs at this pace. Full measurement protocol.
What you configure
Section titled “What you configure”- Providers: each one is declared in the box, with its key.
- Each provider’s safety: a provider is marked safe or not, with its type,
localorZDR/DPA. Only safe providers serve confidential work. - The Local table: the confidential routing table, axis by axis, composed in the box from safe models only.
- Open routing: a setting published by TMB (catalogue
eco,full,eu…), pulled by the box and enabled by the administrator.
Pricing
Section titled “Pricing”On request, for each level.