What changes when I move from my workstation to production?

Not the OS. The kernel, the NVIDIA driver, CUDA, and the containers you built against are the same ones that run in production. Moving to production is a license change, from Developer to Enterprise. What gets added is enterprise support, indemnification, and security response from CIQ. And against Enterprise Linux 9's 5.14 kernel, the same tuned kernel loads models 33.8% faster on identical storage.

Do I have to rebuild my containers when the driver updates?

No. The containers and the host operating system come from the same vendor and are tested together, so a host driver upgrade doesn't require rebuilding the container. And running in the container doesn't cost you throughput: in CIQ's testing, models loaded as fast inside it as on the host.

What does it cost to start?

The RLC Pro AI Developer license is free for individuals: up to 25 nodes for non-production, non-commercial use, self-supported, renewed annually. Production runs on the Enterprise license, priced per node. Schedule an evaluation with our sales engineers to scope your environment.

RLC Pro AI

AI speed with Enterprise Linux stability

Enterprise stability with the newest long-term kernel and validated NVIDIA drivers. Built to keep pace with AI with the same Linux OS from desk to production.

RLC Pro AI benchmarks

How fast is RLC Pro AI?

Up to 10% faster on the NVIDIA DGX Spark and faster model loading than Ubuntu.

Read the benchmark

249
Tokens/s on DGX Spark
Qwen3.5-0.8B, one user sending 2 requests at once, 245 to 249 across runs
3m 44s
Happy-path setup
Node deployment with CUDA, the driver, and PyTorch already working
18%
More model-loading throughput
Against Ubuntu 24.04 with the HWE kernel, cold 8 GiB read from NVMe
0%
Container overhead
Models load as fast inside the container as on the tuned host

Trusted in production by global enterprises for their most advanced AI workloads

The problem

Keep your AI speed when work moves to production

Fast in dev, rebuilt to ship
Dev stacks run new kernels while production Linux runs older ones, so teams rebuild to ship.
Old kernels, new GPUs
Fixed enterprise kernels trail new GPU generations and CUDA releases, and AI work slows down.
Every update is a project
A driver or CUDA update can break a working build, and your team owns the fix.

The solution

One OS from your desk to the data center

Get the newest long-term kernel, validated NVIDIA drivers and AI frameworks like PyTorch and Ollama, then ship your prototype as is to production with expert support.

Features

Built for developers. Supported for production.

Tokens per second on NVIDIA DGX Spark with RLC Pro AI 9.8. Qwen3.5-0.8B: 196 to 199 with a 16K context and 2 requests at once, 122 to 125 for a single user, 245 to 249 for a single user with 2 requests at once. Qwen3.5-2B: 87 to 97 with a 16K context and 2 requests at once, 47 for a single user.

Ranges across 3 to 12 runs, not best runs. Throughput reflects the whole setup: recipe, container, OS, and driver.

Technical specs

What ships in RLC Pro AI

Everything your build depends on, maintained by CIQ. Exact versions are detailed in the individual release notes.

What ships in RLC Pro AI
Kernel	CIQ Linux Kernel, built from the current upstream long-term release.
CUDA toolkit	Current release, validated against the shipped driver and kernel.
NVIDIA drivers	Long-term support branch by default, with a newer branch available.
AI frameworks	PyTorch and Ollama, with more continuously being added.
Container tiers	Base, runtime, devel, workload.
Free to use	Devel container image.
Architectures	x86_64 and aarch64.
Delivery	Boot ISO, KVM, cloud images, OCI containers from the CIQ registry.
Developer license	Free for individuals, up to 25 nodes, non-production.
Marketplaces	AWS, Azure, Google Cloud.
Compatibility	Enterprise Linux compatible.
I/O	io_uring enabled by default.
Tuning	AI-tuned kernel profile for model loading and inference.
Support	Standard and Premium commercial SLAs, with RLC Pro's indemnification and security response.

Backed by the team that builds the kernel

When the prototype becomes production, RLC Pro AI inherits RLC Pro's indemnification and security response.

Same OS for dev and production environments.

Start on your desk. Keep the same OS when you ship.

Frequently asked questions

On a DGX Spark running RLC Pro AI 9.8 with the 595 driver, Qwen3.5-0.8B served 196 to 199 tokens/s with a 16K context and 2 requests at once, and up to 249 tokens/s for a single user sending 2 requests. Qwen3.5-2B served 87 to 97 tokens/s. Each range spans 3 to 12 runs. Throughput reflects the whole setup, so expect yours to vary with the model and configuration.

Resources

View all posts