Evidence before promises

Built, Tested, and Still Being Proven.

Kinode has demonstrated segmented Qwen3-8B inference across physical Apple Silicon and NVIDIA CUDA devices. Implementation alone is not treated as acceptance.

Apple SiliconMetal · coordinator

Model segmentsCoordinated inference

NVIDIA GPUCUDA · worker

Demonstrated physical topology · Qwen3-8B Q4_K_M

Engineering status by capability
Capability	Status
Heterogeneous segmented inference using Apple Silicon and NVIDIA CUDA	StatusDemonstrated
Physical execution of Qwen3-8B in a laboratory topology	StatusDemonstrated
Capability-aware resource planning	StatusImplemented, qualification ongoing
Resource admission	StatusImplemented, qualification ongoing
Owner commitments	StatusImplemented, qualification ongoing
Leases and resource protection	StatusImplemented, qualification ongoing
Oversized-model Capacity Unlock	StatusIn progress, physical proof pending
Sustained performance benchmarking	StatusIn progress, physical proof pending
Reliability qualification	StatusIn progress, physical proof pending
Controlled beta user experience	StatusIn progress, physical proof pending
Managed inference API platform	StatusFuture direction
Permissioned compute supply network	StatusFuture direction
Commercial marketplace and billing	StatusFuture direction

Read the technical boundaries

The demonstrated path partitions Qwen3-8B Q4_K_M between a Mac-based coordinator and an NVIDIA CUDA worker. It establishes real remote participation in a bounded laboratory configuration.

This is not a production-readiness or performance-advantage claim. Timing boundaries differ across existing measurements, so the figures below are not a performance comparison. Repeatable oversized-model execution, broader compatibility and reliability remain qualification milestones.

Historical laboratory measurements
Run	Decode	Context
Distributed path, headline	~1.23 tokens/s	Mac coordinator and remote GTX 1050 CUDA worker. 5 input and 2 output tokens, 3 repetitions, product-level timing.
Distributed path, separate timing view	~0.955 tokens/s	A distinct timing boundary, not interchangeable with the headline.
Single M2 Pro, local	~23.37 tokens/s	Same model family, separate setup and timing conditions.

These are short laboratory observations, not a like-for-like sustained benchmark. The arithmetic ratio of about 19 is illustrative only, and neither a speedup claim nor proof of capacity beyond the best individual node follows. Distribution is not presented as a performance win today. The reason to pursue it is capacity.

The experience taking shape

A clear view of your compute.

Product UI preview

Models, readiness and owner controls, brought into one native experience. Supplied interface previews include illustrative data and planned surfaces; they are not live activity, customer evidence or measured results.