// odyssai
Odyssai
Start your OdyssAI. Earn your sovereignty.
Odyssai is a European local AI ecosystem. A frontier-grade AI experience on infrastructure you control. Not a compromise: the path to technological sovereignty.
Open-source. Apache 2.0. Pre-release, already sailing.
Two products, one stack
// odysseus
The engine
Local inference engine built on Apple MLX and mlx-distributed. Distributed inference across up to five Apple Silicon nodes over Thunderbolt 5 RDMA. OpenAI- and Anthropic-compatible APIs.
// companion
The experience
The agentic layer: conversations, editable persistent memory, projects, files, sessions. Modern interface with VS Code integration and MCP compatibility.
What changes
- Sovereign memory. Persistent, editable, scoped. Not a black box owned by a provider.
- Semantic routing. Every request directed to the optimal model (chat, deep analysis, code) via local embedding.
- 1 to 5 node cluster. From 122B to 1T parameters depending on your configuration.
- Durable sessions. Streaming, cancellation, state recovery. Long-running work actually works.
- Visible tools. Agentic workflows are transparent, permissioned and logged. No magic.
- Multi-model. Local open-weight or cloud models. You choose.
// tmb benchmark
Qwen 3.5 397B BF16 on a 4-node Odysseus cluster reaches 94.4% of Claude Opus performance. The ceiling is no longer the cloud.
"The decision belongs to you, not your client."
fp16 is what the model was trained at. Q4 is a convenience, not a truth. No magic fog. No fake autonomy. The stack stays replaceable through standard protocols.