CoeOS — the "best at everything" virtual model

In one sentence: CoeOS is a single virtual model that sends every request to the model our benchmarks proved best at that specific skill — from Python code to GDPR to long-form writing. Your agents call one model, CoeOS, and get the best of the whole fleet, automatically.

How it works

CoeOS doesn’t run inference itself. It classifies the request, then relays it to the right model:

Agent request  →  CoeOS
                    │  1. Which skill? (Python? GDPR? creative?…)
                    │     → a small "decider" model classifies it into 1 axis
                    │  2. Which model is best at that axis?
                    │     → the TMB Settings table says so (from our benchmarks)
                    ▼
              the winning model (local or cloud)  →  response

Three concepts:

Term	Definition
Skill axis	A benchmarked category: `python`, `react`, `swift`, `legal_rgpd`, `creative`, `debug`, `plan_spec`… 15 axes in total.
Decider	A small, fast model (e.g. NEX-N2-mini) whose only job is to classify each request into one axis.
TMB Settings	The file that defines, per axis, which model to serve — derived from our TMB scoreboards. This is the “recipe”.

Classification follows this order:

Explicit header — if the agent knows its skill, it states it (x-coeos-axis: coding) and CoeOS uses it directly.
Decider — otherwise the decider model classifies the request into an axis.
Default axis — if the decider is unavailable or unsure, CoeOS routes to the default axis (the generalist).

Then: axis → bound model. If that model isn’t warm, CoeOS automatically falls back to another available model (and logs it) instead of failing.

Advantages

The best of the whole fleet, effortlessly. One endpoint; each skill goes to its champion. No per-agent configuration to maintain.
Evidence-based, not marketing. The choices come from our TMB benchmarks (real per-task scores), not a generic leaderboard.
Sovereign or “best of all”, your choice. A local profile (100% on your machines) or a cloud profile (we allow the best online models where they dominate).
No code to change to evolve. The taxonomy and the model choices live in a file. A new model wins an axis? Edit the file.
Transparent. The response reports the model actually used (CoeOS · python — MiniMax-M3).

Use cases

CoeOS is built for agents and orchestrators (Omnigent, Cline, Aider, Hermes…), not for chat (the chat client keeps its own router).

A multi-step agent that plans, codes, debugs, writes: each step goes to the model that scores best on it, without the agent having to know.
A plan → code pipeline: the planning phase goes to the best planner, the coding phase to the best coder.
A heterogeneous fleet (several local + cloud models) you want to exploit optimally behind a single address.

Configuring CoeOS

In the OdyssAI-X dashboard → Settings → CoeOS.

1. Import a TMB Settings file

Easiest path: Import TMB Settings… and pick a shipped file:

Sovereign (local) — everything on your machines.
Best of all (cloud) — allows cloud models where they win.

The import fills everything at once: the 15 axes, the decider, the default axis.

2. Adjust (optional)

Everything stays editable after import:

Axis grid (3 × 5) — one dropdown per skill. Pick the model from every published one (local + cloud). The same model can cover several axes. An axis left empty falls back to another.
Decider — the model that classifies. Small and fast. It must stay loaded (see the note).
Default axis — where to route when in doubt (typically the generalist).

Every change is saved immediately. Export lets you download your profile as a file.

3. Enable

The Enabled toggle. Once on, CoeOS appears in the model list (/v1/models) and in clients as an endpoint, alongside Argo or Telemak.

⚠️ The decider must stay warm. Per-skill routing depends on the decider model: if it’s not loaded, all requests fall to the default axis (no fine-grained routing). Reserve a slot for it — it’s small.

Cold-boot autoload (optional): if an axis’s model isn’t warm, load the default axis’s model anyway instead of returning an error. Off = strict.

Using CoeOS

CoeOS speaks the OpenAI API. Point any agent at it, model = CoeOS.

export OPENAI_BASE_URL="http://<your-odyssai-x>:8000/v1"
export OPENAI_API_KEY="dummy"          # no key required on a LAN

curl "$OPENAI_BASE_URL/chat/completions" \
  -H "Content-Type: application/json" \
  -d '{"model":"CoeOS","messages":[{"role":"user","content":"Write a Python script that…"}]}'

CoeOS classifies (“Python”) and serves the best Python coder in your profile.

Forcing an axis (optional)

If your agent already knows the step’s skill, it can bypass the decider with a header — faster and safer:

  -H "x-coeos-axis: legal_rgpd"

Available axes: creative, legal_rgpd, legal_complex, reasoning, calc, python, code_general, debug, react, swift, refactoring, plan_decompo, plan_spec, plan_judgment, fast_tools.

Seeing which model answered

The response exposes the model actually used. Compatible clients (e.g. Companion) display it: CoeOS · python — MiniMax-M3.

The TMB Settings — the “recipe”

This is the heart of CoeOS, and a deliverable in its own right: a file that states, per skill, which model is best according to our benchmarks. We update it monthly as new benchmarks and new models land. Format:

{
  "name": "TMB Settings — Sovereign (local) · v0.1",
  "regime": "local",
  "decider_model": "telomnis:nex-n2-mini-6bit",
  "default_axis": "code_general",
  "axes": [
    { "key": "python", "label": "Python / scripts", "model": "default",
      "bench": "T03 48.5 / T04 48.5 (M3)", "verified": true },
    { "key": "swift",  "label": "Swift", "model": "",
      "bench": "to benchmark", "verified": false }
  ]
}

empty model = axis not yet filled (waiting for the best benchmarked model).
bench = the evidence behind the choice (the TMB test and the score).
verified = false when the binding is estimated rather than measured by us.

Adding a skill = adding a line. No code to change.

In short


What	A virtual model `CoeOS` that routes per skill.
For whom	Agents / orchestrators (not chat).
How	Decider classifies → TMB Settings names the best model → relay.
Config	Settings → CoeOS: import a TMB Settings, adjust, enable.
Usage	OpenAI endpoint, `model: "CoeOS"`, optional `x-coeos-axis` header.
The edge	The best of the whole fleet, grounded in real benchmarks.