Inference requests are scored on model-cache affinity, prefix-cache match, EWMA-smoothed latency prediction, KV-cache pressure, worker health, utilization, region affinity, per-project fairness, and tier, staleness, and ownership adjustments, then sampled with the A-Res weighted-random algorithm and 5% epsilon-greedy exploration to keep new routes discoverable. Tool execution dispatches through lease-bound agents over a CurveZMQ mesh with MessagePack framing, credit-based streaming backpressure, and an atomic credit ledger.
Six capabilities the platform exposes today. Each links to its documentation.
Drop-in replacement. One env var change ships your app to dedicated, isolated infrastructure -- keep every SDK and library you already use.
Push a model, get an endpoint. Automatic scaling from zero.
Runtime, data, and network isolation.
NVIDIA and AMD support giving full featured acceleration, at scale.
Upload safetensors models with hardware-isolated sandboxing.
Real-time request tracking and performance metrics.
Wire Cursor, Claude Code, opencode, VS Code, or Claude Desktop into your Xerotier workspace once. Then keep working.
Memory, decisions, milestones, artifacts, code review, all reachable from inside your AI tool. No copy-paste. No context reload. One connection per workspace.
Pay per token on managed inference, or bring your own hardware and pay nothing per request. Only tiers with active infrastructure shown.
Perfect for testing and evaluation
ZenDNN-optimized AMD EPYC inference
Best balance of price and performance with NVIDIA CUDA
Cost-effective GPU inference with AMD ROCm
Bring your own accelerators
Free tier, no card. Deploy a shared endpoint, or bring your own hardware whenever you want.
Start free