Kvasir network

A decentralized AI-inference network (DePIN) where everyday devices serve open models and earn KVR.

Kvasir is a decentralized AI-inference network: large open models are split across shared hardware with the linkcpp engine, so no single node holds the whole model. Anyone can contribute a GPU, CPU, NPU — even a phone — and earn KVR for the layers or experts their device actually serves. Developers reach the network through OpenAI/Anthropic-compatible gateways and pay per inference.

Source-available engine — linkcpp is licensed under the BSL (free for development and testing, production use requires a license); the inference engine data plane underneath stays stock and inspectable.
Non-custodial — rewards settle to each node owner's own Solana wallet; keys never leave the user.
Proven on real hardware — a 122B model has run split across 4 AMD MI250 GPUs, across a heterogeneous fleet of GPU/CPU/NPU/mobile nodes, with per-node contribution credited end-to-end.
Named after the Norse myth — Kvasir, the wisest being, born from the pooled essence of every god and owned by none.

One request, many devices

How an inference flows

Every hop is ordinary HTTP/TCP; the model itself is what's distributed.

client SDK ──▶ gateway (OpenAI/Anthropic API, KVR settlement)
        ──▶ hub controller (plan · orchestrate)
        ──▶ serving topology: pipeline ring over layer windows,
            or expert-swarm dispatch at (layer, expert-range) grain
        ──▶ token streams back · each node's contribution is credited

Roles stack: one machine can be a compute node, a gateway host and a hub host at once, and its rewards sum. The network's job is to make the aggregate look like one machine — one endpoint in front, thousands of imperfect devices behind.

Today the network runs on Solana devnet; KVR is a utility / contribution token, not a tradable asset or investment, and nothing on this page is financial advice.