OdyssAI-Xthe inference engine.

Cluster
Overview

The cluster bar at the top shows every logical cluster at a glance — Argo, teleFast, Telemak, Telemak Multi. Click to inspect any one.

Per-cluster: loaded model, node count, uptime, tokens/s, and full node telemetry with RAM, vRAM, and active runners.

Pools &
Nodes

Inside each cluster: named pools group nodes by purpose. The default pool serves interactive inference; a minimax pool runs a different model concurrently.

Node telemetry: RAM, vRAM, load latency, PIDs. Activity badges — Idle · Waiting · Running — update live.

teleFast
& Telemak

teleFast — single-node, native, two models loaded simultaneously. 73 GB wired, 15 tok/s, optimised for throughput.

Telemak — multi-node pipeline. Smaller models, more parallel sessions, tensor-parallel coordination. Telemak Multi extends to multi-modal workloads.

Model
Matrix

The Models tab shows every model across every cluster in one matrix. Rows are models, columns are clusters — quantization and load status at a glance.

Load, unload, or swap quantizations in one click — no SSH, no shell commands. See redundancy for failover planning.

HuggingFace
Downloads

Pick target clusters and nodes, paste the HF repo ID, authenticate with a saved token, and start. Downloads run in parallel via hf download across all selected nodes simultaneously.

Progress streams to the UI in real time. Once complete, the model appears in the matrix and is ready to load onto a pool.

CoeOS

CoeOS classifies each incoming request by skill — conversation, code, analysis, expert — and routes it to the configured cluster and pool.

A waterfall plan defines fallback order: local cluster first, then secondary, then cloud. Saturated? Reroutes silently in <1ms.

This is what powers Companion's Easy mode — automatic model selection, invisible to the user.

One URL.
Any client.

OdyssAI-X exposes a single OpenAI-compatible base URL. Pick a model alias, copy the config snippet, paste it into your client.

Ready-to-paste configs for: OpenAI SDK · Anthropic SDK · Continue.dev · Hermes. No proxy — your cluster, your model, your data.

Cloud
Fallback

Register OpenAI, Mistral, Ollama, or any OpenAI-compatible provider. CoeOS places them in the waterfall wherever you need them.

Local cluster saturated? Fall to Mistral. Mistral errors? Fall to OpenAI. The user sees uninterrupted responses. The endpoint format is identical across all providers.