European · open-source · local-first

The best open-weight models. No compromise.

A complete European AI ecosystem: an inference engine for Apple Silicon clusters, and an AI operating system on top. Local when it matters. Cloud when you choose. Your models, your data, your hardware.

See the problem→ Explore the stack Download Nemo ↗

/01 · Problem

Adopting AI in the enterprise means giving something up.

Option A · Proprietary cloud

Polish, no sovereignty.

Quality, modern agents, the user experience of a real AI OS. But your data leaves for the vendor, you're locked to their catalog, and every request is one more GDPR liability.

Option B · Local AI as it stands

Sovereignty, no polish.

Your hardware, your weights. But too often a chat box and nothing else: no real memory, no agents, models capped by a single machine's RAM. A trip back to 2023.

Between data leakage and local tinkering, a third way was missing.

/02 · Solution

OdyssAI removes the tradeoff.

Frontier-grade AI experience, on infrastructure fully under control. Sovereignty and capability in one system. Local when it matters. Cloud when you choose.

Pillar 01

Sovereign by design.

Inference, routing and memory live on hardware you own. Prompts, files and keys never leave your network unless you bound that request to a cloud model yourself.

Pillar 02

Frontier-grade, local.

An Apple Silicon cluster runs the largest open-weight models — Kimi, GLM, DeepSeek, MiniMax, Qwen, Mistral, … — in their best quantization.

Pillar 03

A complete AI OS.

CoeOS: a desktop client, a router that sends each request to the model proven best at that task, and the decision shown to you every time. An interface you can live in.

/03 · Hybrid

Frontier cloud when it helps. Local when the data demands it.

OdyssAI is hybrid by design. At its heart, the CoeOS box routes every request to the model proven best at that task, cloud or local, and Guardian flags anything confidential before it leaves.

97.2/100

CoeOS v1.3, the OdyssAI router — statistically on par with Fable 5 (96.9) and Claude Opus 4.8 (96.0) on equal-footing API runs. The spread is within stochastic noise; the cost isn't: 4.9× less per test than Fable 5. The best of open models, in one endpoint.

See the full scoreboard→ Meet CoeOS → CoeOS docs →

Routing by skill: every request is classified on a competence axis — by an explicit header from the client, or by a small decider model you configure. Writing, law, analysis, code: the strongest model on 18 measured competency axes takes the turn, and the answer names it.

Guardian: a request can be screened before it leaves. Personal data, client files, trade secrets are detected; the client then decides — warn, keep it on a local model, or block. Guardian detects. It never decides for you.

A stated engineering position: no model is served below Q6, or Q4 when it was trained specifically for that quantization (Kimi, for example). On TMB Benches, GLM 5.2, Hy3 and MiniMax served locally show no significant loss of accuracy against their cloud versions.

GLM 5.2 Cloud (480/500, 19 June 2026) beats Opus 4.7 on the T01 benchmark. See the general scoreboard →

/04 · Stack

Two pieces. One sovereign stack.

One engine that runs the cluster. One AI operating system that routes, remembers, and shows you which model answered. Both built around standard wire protocols so you can plug your own clients in.

04.a · Cluster engine
OdyssAI-X
carries the weight.

Distributed inference and orchestration on Apple Silicon. Split a model across nodes when it does not fit one Mac, or run one full copy per Mac and serve a crowd. The largest open-weight models in their best quantization, built on MLX, over RDMA on Thunderbolt 5 or plain TCP.

For operators and teams who want frontier-class compute without giving up standard clients.

Surface OpenAI · Anthropic
Ground Apple Silicon cluster
04.b · AI operating system
CoeOS
gives the experience.

Nemo, the desktop client: chat, cowork on documents, code with a panel of agents, a memory kept in a vault on your Mac. The CoeOS box, the router: one address for every tool you own, each request sent to the model proven best at that skill, local or cloud, with your keys.

Every answer names the model that wrote it. Nothing leaves your network except the request you chose to send out.

Surface macOS app · OpenAI · Anthropic
Holds Routing · Memory · Guardian

Also A sharpened, dedicated version for coding is available as CodeOS: a modified opencode tailored with our agentic workflow, where separate models review and validate the plan and the code. CodeOS on GitHub ↗

Professions For professions that cannot send a file to the cloud, CoeOS also comes as a paid suite: the same client and router on a dedicated server per client, with Guardian customised to your profession's confidential categories on top of GDPR. Talk to us →

/05 · Experience

Local AI should not feel like a downgrade.

You shouldn't have to choose between privacy and polish, between local models and modern agents.

→ 01

Memory that stays on your Mac.

A personal memory in a local Markdown vault, Obsidian-compatible, injected at startup. Project and company corpora when a box provides them. Not a black box owned by a provider.

— personal
→ 02

Models you can choose.

Open-weight local models when sovereignty matters. Cloud models when you decide they're the right tool. Routing is per skill; the catalog is yours.

— optional
→ 03

Routing you can see.

Every answer names the axis it was classified on and the model that served it — in the picker, and in the response headers. No hidden model behind the curtain.

— transparent
→ 04

Tools without theatre.

Agentic workflows that are visible, permissioned, interruptible, logged. No magic fog. No fake autonomy.

— visible

/06 · Internals

Three mechanisms that make the difference.

The honest version. Open a block if you want to know how. Skip if you just want to know it works.

I.01 How a request is routed.

Your tools call one model, CoeOS. The box classifies the request on a competence axis in three steps: an explicit x-coeos-axis header from a client that already knows its phase; otherwise a small decider model you configured; otherwise the default axis. Then axis → the model your TMB Settings bind to it, on your engine or at a cloud provider.

The routing table is data, not opinion: it comes from the TMB benchmark, scored axis by axis, and you can read and edit every binding. A new model wins an axis? Change one line.

The answer comes back untouched, in the caller's own format, with x-coeos-axis, x-coeos-model and x-coeos-provider in its headers. You always know who spoke.

I.02 Guardian, two stages.

Stage one is a privacy model — GLiNER2 — that finds identifiers and GDPR-style identification in about forty milliseconds, always on. Stage two, opt-in, asks a small local instruct model the question named-entity recognition cannot answer: is this a business secret, a health narrative, an HR or legal matter?

The rule is short. A salary figure alone is a topic. A salary figure tied to a name is personal data. sensitive = hard_identifier OR secret OR (identity AND attribute).

Guardian returns what it found. What to do — warn, keep the request on a local model, block — is the client's policy. Detection and decision are kept apart on purpose.

I.03 Replica mode.

Sharding makes a model fit. Replicas make it serve a crowd. When the model already fits one Mac and what you need is users, not size, OdyssAI-X loads one full copy per node and hands the pool to your clients as one model name.

Least-busy dispatch. Session affinity: a conversation stays on the replica that holds its KV cache, so nothing is re-prefilled. Optional continuous batching: a node serves several conversations in the same forward pass instead of one after the other.

Self-healing: a replica that dies — crash, memory kill, reboot — is restarted by the pool and put back into dispatch. No inter-node collective, so no node cap. Add a Mac, gain throughput.

/07 · Europe

Sovereign infrastructure for a sovereign continent.

Data residency is not a setting. There is no cloud to reside in.

Built in Europe, for European regulatory reality. Fully compatible with Mistral models, alongside the best open-weight models worldwide. A European frontier model, on European hardware, end to end.

/08 · Status

Pre-release. Running in production, every day.

The only question an enterprise truly asks: will it still be running on Monday? Yes.

OdyssAI-X carries our own work on a rack of Mac Studios — several clusters side by side, distributed, replica and vision pools at once. The engine is AGPL-3.0; Nemo, the CoeOS box and Guardian are MIT. The current cycle is about putting it on someone else's hardware.

OdyRAG

Under the hood

OdyRAG — our RAG.An improved LightRAG with an easy-to-use web UI.

OdyRAG docs →

/09 · Contact

Let's talk.

Cluster deployments, the CoeOS suite for your profession, partnerships, integrations, press, or simple technical curiosity. Answered from Europe, in French, English, or Spanish.

Send a message→