TokenX single-node inference is ready to use and verify. Read the technical overview

Multi-GPU Inference for Distributed Compute

TokenX is built for multi-GPU, heterogeneous, and distributed compute environments. It coordinates AI inference, optimizes performance, and manages resources to help make better use of available GPU capacity.

Products

Inference orchestration
for heterogeneous compute

Single-node inference loop A verifiable path from request to settlement

TokenX provides a unified inference endpoint and a verifiable single-node workflow spanning request routing, usage metering, and settlement records.

OpenAI-compatible inference API

Submit chat requests to /v1/chat/completions with stream=true for streaming responses. Available models and node capabilities depend on admission status.

View details

Single-node inference loop

Authentication, five-stage routing, vLLM / oMLX inference, streaming, tiered billing, and idempotent settlement form the currently testable single-node workflow.

View details

Reliability and budget controls

Disconnect cancellation, rate limits, node timeouts, budget limits, and structured request logs provide the currently implemented request safeguards.

View details

Built for real-world inference
workloads

TokenX is designed for multi-GPU, heterogeneous, and distributed compute. Model partitioning and cross-device coordination let different devices work together on large-model inference while continuously improving end-to-end efficiency.

Run one large model across distributed GPUs

Explore capability boundaries
Use real device measurements to evaluate compute, batching, concurrency, and cross-device handoff overhead—and verify whether optimizations actually improve aggregate output throughput.
Optimize end-to-end inference throughput
Evaluate feasible model-partitioning strategies across GPU models, memory capacities, and device counts instead of relying on a fixed hardware configuration.
Adapt to different device configurations

Partition a model across multiple GPUs so large models that cannot fit on one device can run using pooled memory across devices.