KINODEGet in touch

Technology

Your computers are more capable together.

More devices do not automatically mean more capacity. What can run depends on the workload and the hardware, so Kinode starts by asking whether to distribute at all.

What computers can run together, by workload scenario
Workload ScenarioSingle-Device ApproachKinode Execution PossibilityMaturity / Evidence
Model fits locallySingle-Device ApproachRun locally.Kinode Execution PossibilityPrefer local execution when distributing offers no useful advantage.Maturity / EvidencePlan selection in qualification
Model exceeds individual memory limitsSingle-Device ApproachCannot fit within its eligible memory budget.Kinode Execution PossibilityEvaluate partitioning across compatible devices, subject to actual weights, KV cache, memory limits, hardware compatibility, network performance and workload requirements.Maturity / EvidenceOversized-model physical proof pending
Multiple independent inference requestsSingle-Device ApproachLimited to one node’s capacity.Kinode Execution PossibilityPotential request distribution across eligible nodes.Maturity / EvidenceProduction-scale qualification pending
Complex task with independent subtasksSingle-Device ApproachSequential execution is one option.Kinode Execution PossibilityPotential parallel subtasks with result synthesis.Maturity / EvidenceResearch / future execution strategy

A model that exceeds one device does not automatically fit across the pool. Partitioning depends on actual weights, KV cache, memory limits, hardware compatibility, network performance and workload requirements.

Execution strategies

One Fabric. Different Ways to Execute.

Select a workload scenario
WorkloadModel stage 1Model stage 2Model stage 3Request ARequest BRequest CSynthesisABCD
Illustrative execution strategy, not live benchmark data.

Recommended approach

Capacity Mode

Potential benefit
Could make a workload possible that no participating device can hold alone.
Important limitations
Not automatic. Depends on actual weights, KV cache, memory limits, hardware compatibility, network performance and workload requirements, and it can be slower than a capable single device.
Capability maturity
Small-model segmented execution demonstrated. Oversized-model physical proof pending.

The fabric

One fabric.
Different devices.

Kinode connects computing resources across different devices to run AI workloads through a trusted, heterogeneous execution fabric.

MApple SiliconMetal
NVIDIACUDA
Compatible backendsCPU / Vulkan*
Device capabilitiesAvailable memoryNetwork conditionsOwner commitments

*Backend and workload compatibility are subject to qualification. Connected does not automatically mean eligible.

How it works

From individual machines
to coordinated execution.

Conceptual diagram: trusted devices connect, owners set limits, and an eligible plan uses only some of them.
Conceptual · trusted devices connectConceptual · owners choose what Kinode may useConceptual · connected does not mean eligible
  1. 01

    Connect

    Discover and connect compatible trusted devices.

  2. 02

    Allocate

    Device owners choose which resources they allow Kinode to use.

  3. 03

    Execute

    Kinode selects an eligible execution plan and coordinates AI inference across participating resources.

NextHow Kinode compares Talk to the Founder