Compare
Not All Compute Platforms Work the Same Way.
Published capabilities and Kinode’s current status, side by side. Real differences, stated plainly.
- exoFocuses on model partitioning and distributed inference across your own devices.
- SaladPrimarily commercializes distributed compute supply and inference services.
- KinodeAims to combine cross-backend model execution, owner-authorized resource participation and workload-aware execution decisions.
Kinode’s combination remains a differentiation thesis, not a proven market advantage. No capability is claimed as exclusive to Kinode.
The core idea
More than connecting computers. Making their resources work together.
Traditional compute platforms typically schedule workloads onto individual machines. Kinode is exploring how heterogeneous devices can also participate in a single model execution, potentially unlocking workloads that cannot fit on any participating device alone.
Each workload must fit on one machine. A larger model finds no place.
Model stages could span compatible devices. Oversized-model physical proof is pending.
Conceptual comparison, not measured results.
Where Kinode fits
exo, Salad and Kinode
exo already proves model splitting. Salad already proves scalable idle GPU supply and private deployment. Kinode has to prove its value between those strengths.
| Capability | exo | Salad | Kinode |
|---|---|---|---|
| Core product purpose | exoLocal distributed model inference and model partitioning. | SaladDistributed GPU and container supply, managed model APIs and private inference. | KinodeOwner-controlled heterogeneous inference fabric. |
| Model partitioning | exoDocumented capabilityCore capability, with automatic splitting across devices. | SaladNot documented as a standard featureSchedules containers and replicas onto eligible GPUs. | KinodeDemonstrated in prototypeSmall model across Apple Silicon and NVIDIA CUDA. Oversized-model unlock not yet proven. |
| Heterogeneous hardware support | exoDocumented capabilityApple Silicon GPUs through MLX. CPU on Linux, with Linux GPU support under development. | SaladDocumented capabilityMainly NVIDIA RTX-class GPUs, with selected AMD classes. | KinodeDemonstrated in prototypeMixed Metal and CUDA participation. CPU and Vulkan remain subject to qualification. |
| Device discovery | exoDocumented capabilityDevices on the same network discover each other automatically. | SaladDocumented capabilityProviders join through Salad’s desktop app, and GPUs are tested before joining. | KinodeIn developmentDevice discovery and capability-aware eligibility, implemented in part. |
| Scheduling and placement | exoDocumented capabilitySplits models using device resources and link latency and bandwidth. | SaladDocumented capabilityFills replica requests from compatible nodes and moves workloads when a node drops. | KinodeIn developmentCapability-aware planning and admission. Sustained performance-based selection needs validation. |
| Resource-owner governance | exoNot documented as a standard featureNot described as a multi-owner resource-leasing product. | SaladDocumented capabilityIdle detection, primary-user priority, contribution windows and IT controls. | KinodeIn developmentExplicit commitments, admission, leases and resource protection. Physical qualification ongoing. |
| Managed inference APIs | exoNot documented as a standard featureOffers local OpenAI, Claude and Ollama-compatible endpoints rather than a managed service. | SaladDocumented capabilityAI Gateway and Private Cloud OpenAI-compatible endpoints. | KinodeIn developmentGateway foundation exists. Managed, metered model serving is not launched. |
| Private deployment | exoDocumented capabilityRuns locally on your own devices. | SaladDocumented capabilityPrivate Cloud keeps prompts and model weights inside the organization’s network. | KinodeIn developmentTrusted groups and controlled networks first. No confidentiality claim against an untrusted host. |
| Public compute marketplace | exoNot documented as a standard featureNot its documented core offering. | SaladDocumented capabilityCommunity Cloud: an on-demand pool of idle consumer GPUs. | KinodeFuture visionSupply, security, payments and service levels are unvalidated. |
| Current product maturity | exoPublic project maintained by exo labs. | SaladProduction services with paid compute. | KinodeEarly prototype. Private beta in development. |
Compares published documentation, checked 4 November 2026, with Kinode’s prototype and roadmap. “Not documented” does not mean a product cannot do it. This is not a feature or performance audit.
Choosing honestly
When something else may serve you better.
- Local inference
- Your model fits comfortably on a machine you already have. One device avoids network and coordination cost entirely.
- Cloud GPU rental
- You need dependable capacity now, have no suitable hardware, and your data can leave your own environment.
- Other distributed runtimes
- You run a cluster of your own devices today. exo and llama.cpp RPC are available now and do not need Kinode’s owner-governance model.
- Managed distributed GPU
- You want production endpoints or idle-GPU pooling now. Salad already operates these services.
Public documentation: exo documentation · Salad products · Salad Container Engine · Salad AI Gateway · Salad Private Cloud · llama.cpp