Technology

Radium runs the whole stack, from the API down to the server firmware.

The platform and the infrastructure are built by one company and tuned to the hardware underneath, so the speed, cost and reliability of every request are set by the whole stack working together.

01

Platform

Hal, Clarke and Tycho each run as a complete system with an API, a router and a pool of models of its own, and each one sits directly on Infra below it.

HalCoding, agents, tool calling and complex reasoning.$2.00 / $10.00
ClarkeRetrieval, RAG and chat.$0.75 / $4.50
TychoClassification, extraction and high-volume work.$0.10 / $1.50

Per million tokens, input / output

Every model speaks both the OpenAI and the Anthropic APIs, so switching is a change to the base URL and the model name, and the SDKs in your code stay where they are.

https://api.radium.cloud/v1

Each model has a router of its own paired with a reinforcement learning engine, and together they pick which model in the pool serves each request. This engine is separate from the one in Infra below.

Behind each model sits a pool of open-weight models that Radium's model curation team selects and tunes, and every model keeps a pool of its own.

02

Infra

Infra joins every data center site into a single pool of compute, so a request runs wherever there is capacity and every site behaves as part of one machine.

Compute is distributed across all of the data centers with no virtual machines between a request and the GPU that serves it, running on Radium's modified server firmware and its own switch software.

A reinforcement learning engine places and scales every request across every site, which keeps FLOP utilization linear as more sites join.

Radium runs modified firmware on every GPU server and its own software on every network switch, so power, cooling and BIOS answer to the scheduler and full InfiniBand throughput reaches every job with no virtualization in the path.

03

Hardware

Radium runs on racks of GPU servers in data centers across several sites, and Radium's network engineering and infrastructure supply teams look after every one of them.

Each data center site is joined to the others by Infra, and the four drawn here are illustrative.

Each site holds racks of GPU servers, and the firmware and switch software that run them sit one layer up in Infra.

The whole stack

One stack, run by one company.

Building every layer from the firmware up is how Radium sets the speed, the cost and the reliability of each request, and your prompts and outputs never train a model at any layer of it.

Independent Benchmarks

Third-party validated. Real workloads, not microbenchmarks.

2x
Lower inference
latency vs AWS

Training & Inference
Under Mixed Load

Benchmarked at

CMU ran a variational autoencoder on concurrent CPU+GPU workloads.

On AWS, interrupt and scheduling noise caused latency spikes as virtualization layers compounded.

On Radium: stable, fast, predictable. Same hardware class, comparable price point.

30%
Higher MFU on
identical JAX|
workloads vs GCP

Model FLOP
Utilization vs GCP

Benchmarked at

Stanford's Center for Research on Foundation Models ran GPT2-XL training on the same software stack, same hardware generation, across both clouds.

Radium consistently reached 50–55% MFU. GCP plateaued at 39–45%.

Radium also outperformed Stanford's own on-premises HPC cluster.

58%
Better utilization at
16-layer depth vs
AWS

Performance
Advantage Scales
With Model Depth

Benchmarked at

At 4 layers: 30% advantage. At 16 layers: 58% advantage. The performance gap compounds as models get deeper — more kernel launches, more memory movement, more synchronization.

Radium's optimizations handle exactly these operations better than a generic virtualized environment.

One line of code to switch.
A different class of performance.

Every new account gets $25 in credit. No card required.

Get your API key
Dense cluster of white dots scattered against a dark background, resembling stars or particles.