CIQ is the software company behind Rocky Linux, an enterprise-grade Linux distribution built as a free Red Hat (RHEL) alternative. The company this week noted that Arcee AI, a frontier model lab known for its Trinity family of models, has selected Fuzzball (CIQ’s platform designed for High-Performance Computing and AI workload orchestration) as its orchestrator of choice.
In other words, Arcee AI will use CIQ Fuzzball to support workload orchestration for the development of its next flagship large language model.
Hefty heterogeneous happiness
Arcee says it will use Fuzzball to coordinate and place selected “compute-intensive workloads” across heterogeneous infrastructure while continuing to use its internally developed training and inference stack.
Fuzzball provides an orchestration layer for scheduling and coordinating workloads across cloud, GPU-provider, and on-premises environments. Arcee’s model-development systems remain independent of the underlying compute provider, while Fuzzball helps the team manage where and how selected jobs are executed. Arcee retains control of its training and inference software, models, data, and deployment architecture.
What makes Fuzzball so fuzzy?
CIQ doesn’t tell us too much about Fuzzball, but we can see that this is not just orchestration i.e. this is the wider realm of “unified compute orchestration” and that’s a step up because it is capable of bridging traditional High-Performance Computing (like scientific simulations) with modern AI workloads (like training large language models or running inference endpoints).
Fuzzball also enables software engineering teams to manage, scale and schedule containerised AI workflows across on-premises GPU clusters, bare-metal hardware, or multi-cloud environments without altering the underlying code.
It is thought that CIQ co-founder Robert Adolph is a fan of astrophysics and the name is a reference to string theory, where a “fuzzball” is a theoretical description of a black hole that replaces a singular point of infinite density with a warm, fuzzy ball of fundamental strings
Don’t re-write the training & inference stack
“Arcee has built the technical systems required to develop and operate models at significant scale. Fuzzball’s role is to give that team a consistent orchestration layer across different compute environments without requiring its training and inference stack to be rewritten for each provider,” said Bjorn Hovland, president of CIQ.
Arcee’s selection may demonstrate how Fuzzball’s HPC orchestration capabilities can be applied to AI workloads operating across heterogeneous compute environments. Fuzzball coordinates workload execution, while model builders retain their own training, evaluation, and inference systems. Teams with strong internal infrastructure gain a portable orchestration layer without replacing the systems they have already built.
“We have built our own training and inference stack because control over the full model lifecycle is fundamental to how Arcee operates. Fuzzball gives us an additional orchestration layer for coordinating workloads across heterogeneous compute environments without replacing the systems our team has built,” said Lucas Atkins, chief technology officer of Arcee AI.
Enterprises, AI labs & infrastructure providers
Arcee’s models are designed to be customisable and deployable wherever software engineers need them, and Fuzzball extends that principle to the layer underneath. Over time, the two companies say that they will collaborate to help enterprises, AI labs, and infrastructure providers deploy models on portable infrastructure they control.
Arcee brings model development, adaptation, and the training and inference systems it builds and operates itself. CIQ brings Linux, cluster-management, and workload-orchestration technologies that can help those systems run across a range of compute environments.
The same architecture gives GPU providers a way to move up the value chain, turning raw compute capacity into a differentiated Al platform with multi-tenancy, portability, and lifecycle orchestration.