Open · Hybrid · Local

Frontier-grade AI, at open-source prices.

Every request goes to the open model with the best score for the task. Confidential work stays where you decided.

An equivalent score at 3.5% of the price of Fable 5.

96.5/100 CoeOS $0.0151 per test
97.8/100 Fable 5 $0.426 per test
97.0/100 Claude Opus 4.8 $0.141 per test

TMB benchmark, 32 tests. Scoreboard of 27 September 2026. The TMB scoreboard →

/01 · Skills

The specialist for every skill.

A closed model is a generalist, billed for everything. CoeOS classifies each request on one of 18 measured skills and sends it, whole, to the model with the best score there.

  • Business writing
  • GDPR and compliance
  • Multi-jurisdiction law
  • Reasoning
  • Calculation
  • Python
  • Debugging
  • React
  • Swift
  • Refactoring
  • Task decomposition
  • Specs
  • Planning judgment
  • Agent execution
  • Agent safety
  • Creative writing
  • Fast tools
  • Voice

Every answer names the model that wrote it.

/02 · Levels

One CoeOS, three levels.

The same client, the same router. What changes: where confidential work goes, and what stays in the cloud.

All in the cloud

Open

Confidential work
No distinction, or European providers under DPA and ZDR. Your choice.
Everything else
The best model per skill, through OpenRouter or European providers.
On your side
Nothing to install.
Confidential stays home

Hybrid

Confidential work
To a model hosted on your premises, on an Apple Silicon server.
Everything else
Keeps the power of the cloud, best model per skill.
On your side
One Apple Silicon server.
Nothing leaves

Local

Confidential work
Stays on site.
Everything else
Stays on site, routed across your local models as in the cloud.
On your side
An Apple Silicon cluster, sized to your teams.

Everything is a setting: providers, contracts, what runs at home.

/03 · Confidentiality

Confidentiality is an architecture, not a promise.

Nothing retained, nothing used to train a model, no leak.

Because the confidential path leads nowhere else.

→ 01

Confidential by default

Every conversation carries a class. Confidential is the starting point; a conversation is opened on purpose, never by accident.

→ 02

No way back

A conversation that went confidential never returns to an open model. If the confidential target is down, the turn fails instead of leaking.

→ 03

One way out

The client knows no provider address and no key. Everything goes through the router, which only sends confidential work to models you marked safe.

→ 04

Memory kept apart

What was said in confidence is stored apart. An open conversation never reads it, and a test checks that continuously.

/04 · Open models

Can you trust an open model?

Yes, when you know who hosts it.

01

A model is not a program.

Open weights are numbers (safetensors), not code. A model sends nothing anywhere; only the software around it can, and in CoeOS that is the router.

02

What matters: who hosts it, under which contract.

The same model runs at a European host under a DPA, or on your own machines. You choose, provider by provider.

03

Measured before it routes.

Every model goes through The Monocle Bear's TMB benchmark: 32 tests, 17 axes, public results.

/05 · Desktop

Two ways to work, one router.

In the browser

CoeOS

Chat, work on documents, review agents, team memory and company knowledge. Nothing to install, for the whole organisation.

BrowserGuarded multi-agent

The agentic harness

CoeOS Desktop

An agent with the whole machine: your files, your shell. It plans, delegates to sub-agents in parallel, and asks for your approval on every command, or no longer does if you decide so. Your Claude Code skills work as they are.

macOS, Windows, LinuxFull machine access

/06 · Local

Everything on site. No compromise.

The Local level runs on OdyssAI-X, our inference engine for Apple Silicon. Three models, three jobs.

94.1 TMB, local Q8

GLM 5.3

Precision

Planning, calculation, writing: when a mistake costs more than a second.

300+ tok/s aggregate

MiMo V2.6 Flash

Speed

Everyday work: summaries, writing, tool calls.

600 tok/s aggregate

Qwen3 Next

Scale

Fully local agentic flows and large teams.

Aggregate throughput: 96 concurrent requests on three Mac Studio M3 Ultra with 256 GB.

The OdyssAI-X engine →

/07 · Professions

Agents that read with your doctrine.

Law, HR, health: agents tuned to a profession, on a dedicated server per client. The agent proposes, you decide.

→ 01

Contract review

The contract is read against the firm's Golden Rules, versioned and auditable. The output: a conformity table rule by rule, with the clause quoted, the page, a severity and a recommendation.

→ 02

HR-policy review

An HR policy checked against the applicable collective agreement. Same table, same evidence, plus alerts when a clause drifts out of bounds.

/08 · Demo

Let's see CoeOS on your cases.

A demonstration on your documents, at the level that fits you. Pricing on request.

Request a demo hello@odyssai.eu