Stardock’s Clairvoyance Adds Hivemind to Share Idle GPUs for Local AI

Stardock’s Clairvoyance is adding a way for teams to put idle computers to work on AI inference. Its Locama local-model manager includes Hivemind, a feature that shares local inference capacity across a company’s machines, alongside a benchmark intended to show how well particular hardware runs AI.

The November 5 release positions Clairvoyance 0.90 as a desktop environment for using local and cloud models together. Locama manages the installation, execution, and orchestration of local models on Windows, macOS, and Linux; Hivemind extends that idea from one computer to a team’s available machines.

Stardock is targeting routine, high-volume business work in areas such as procurement, supply chains, legal operations, and finance. Its argument is that some of those tasks can run on hardware an organization already owns, reducing the need to send every request to a metered cloud service.

Hivemind Turns Spare Computers Into Shared Inference Capacity

Hivemind allows team members to run AI work using idle GPUs across the company’s computers. That could be especially relevant outside normal working hours, when machines capable of running local models may otherwise be unused.

“Most enterprises have capable machines sitting idle every night, and Hivemind lets them pool those machines and put them to work,” said Brad Wardell, CEO of Stardock Software.

The company’s technical terminology describes Hivemind as shared local inference among peers. It separately identifies Locama as the inference host and a routing component, called Concierge, as responsible for choosing a model and machine for a turn.

That is an important distinction from assuming several computers become one large accelerator. Sharing work across available machines can increase usable capacity, but the announcement does not specify tensor-parallel execution, combined GPU memory, or how one oversized model would be divided between peers.

For a team deciding where to run a workload, the useful questions are whether the model fits the available hardware, whether it produces acceptable results, and whether routing work to another machine adds more delay than it saves. Hivemind’s value will depend on that combination, rather than GPU availability alone.

Local and Cloud Models in the Same Workspace

Clairvoyance organizes AI assistants as named Staff that work with a user’s files and produce documents and other artifacts. Its documentation describes assigning different providers or local models to Staff members, allowing a workspace to mix computing options.

The distinction between those options matters. A local model executes on hardware controlled by the user or organization. Connecting a cloud provider’s tool gives access through that provider’s account and authentication; installing Clairvoyance does not automatically include unlimited access to every commercial model it supports.

The documentation also notes that tool-use capabilities vary among local models. A model that can answer questions is not necessarily able to execute the agentic workflow a team needs. Choosing a smaller local model is therefore a task-specific decision involving capability as well as speed.

That gives the product a practical role between an employee and the underlying model runtimes. Rather than making every task use the same backend, the environment can support different choices for different work. The release does not provide independently measured cloud-cost reductions from Hivemind deployments.

A Benchmark for the Hardware People Actually Own

Locama also includes the Clairvoyance Local AI Benchmark. Stardock describes its headline measure, Model Speed, as combining the time a local model takes to process a request and return a complete response. Public results are available at clairvoyanceai.com/benchmark, and Professional users can run tests on their own hardware.

The product’s release notes describe measurements including prefill, decoding, time to first token, and memory, together with task-oriented checks. Prefill is the processing of an input prompt and its context; decoding is the generation of the response. Those stages can stress hardware differently.

A fast first token can make an interface feel responsive, while the total completion time matters for a batch job or a longer agent task. Keeping those measures visible makes a benchmark more useful than relying on a single hardware specification.

The results page identifies submissions as coming from Clairvoyance users and reflecting their hardware and settings. That makes the database a source of configuration-specific evidence, rather than a controlled independent ranking. Meaningful comparisons still need to account for the model, precision, context, task, and runtime used.

Qualcomm Integration Adds Another Local Compute Option

Clairvoyance’s work with Qualcomm predates this release. A September 23 Qualcomm developer article documents its integration of GenieX, the chipmaker’s on-device generative AI runtime, with the Hexagon NPU on Snapdragon X Series PCs.

That integration broadens local inference beyond conventional desktop GPUs. Qualcomm describes selecting models and compute resources according to task requirements, with support for local agentic workflows. It does not establish that every local model can replace a frontier cloud model for every task.

The relevance to Hivemind is the broader hardware strategy: make existing computing resources usable through one working environment. Different machines may offer different model capacity, bandwidth, and acceleration, so software must match the job to what is available.

Pricing and the Economics of Reusing Hardware

Clairvoyance 0.90 is being released for Windows, macOS, and Linux. Its pricing page lists a free tier for personal and non-commercial use, while Professional costs $20 per month and adds commercial licensing, Hivemind, team-oriented capabilities, cloud sync, and an included AI token package. Enterprise is listed at $100 per seat per month, with additional organizational features and volume arrangements.

Local inference can reduce usage-based cloud spending, but it still carries costs: power, hardware wear, administration, and the opportunity cost of occupying a machine. A cost comparison also needs to include the quality of the output and any work required to review or correct it.

Stardock’s bet is that enough recurring work can meet those requirements on hardware already sitting nearby. Hivemind gives teams a way to explore that proposition, while the benchmark provides a starting point for measuring which models and machines are suited to the job.