DePIN · Frontier models without the datacenter

You don’t have the hardware for a 198B model. Neither do we.

Nothing in this fleet can hold Step-3.7-Flash — 198 billion parameters, far past any single device we own. Kvasir cuts it into layer windows so each machine holds only its own and hands a hidden state to the next. Measured on the serving ring today: 6.7–6.8 tokens a second, first token in 70 ms.

  • Runs on GPU · CPU · NPU · phone
  • OpenAI + Anthropic compatible
  • Source-available (BSL 1.1)
  • Solana devnet

AI shouldn’t be owned by a handful of companies

Frontier inference is concentrating behind a few walled datacenters — closed weights, metered access, one bill paid to one owner. Kvasir is built the other way: open models served across a network of everyday devices, owned and earned by the people who run it. Today that engine and its settlement run on our own fleet, and a device can already register, contribute and be paid; opening the serving ring itself to anyone is the next gate. The roadmap below says where each part stands.

  • Any device joins a peer-to-peer ring — no central master in the ring
  • Source-available p4 engine — BSL 1.1 and fully inspectable
  • Contributors earn KVR for the real compute they give
  • Self-custody wallet — your keys never leave your device

One model, many devices, paid per layer

No single node has to hold the whole model. A request flows across the layer path and each node is rewarded for exactly the work it did.

  • Split: The model is divided into contiguous layer windows. Every device keeps a copy of the model file but loads only its own window into memory, so no node has to run the whole thing.
  • Serve: The request enters the ring. Each node runs its layers and passes only the hidden-state boundary to its neighbor; the last node samples the token and sends it back around — no central master.
  • Reward: Each node earns KVR weighted by the layers it ran — its share of the tokens produced — settled on Solana to that node’s own wallet.

p4 — the engine behind the network

p4 binds mixed hardware into one serving pool: each device holds a few of the model's layers and passes only the hidden state to its neighbour, with no central master in the ring. It supersedes linkcpp, the engine Kvasir ran through mid-2026, and keeps the data plane close to upstream with a small patch set.

  • Ring runtime: Every device stores the same model and loads only its layer window, then opens one link to its predecessor and one to its successor. Hidden-state boundaries circulate around the ring and the last rank returns the token — no central master, no node holds it all.
  • p4 control plane: The control plane the RPC data plane was missing. It discovers devices, plans layer and expert placement, launches the workers, and exposes the gateways. A placement plan is an operator artifact — the engine will not load a model because a web request asked it to. Source-available under the Business Source License (BSL) 1.1.
  • Distributed layer placement: p4 reads GGUF metadata and computes contiguous per-node layer windows from a placement plan, plus optional MoE expert-FFN offload to node RAM. Step-3.7-Flash, a 198B MoE, currently sits as 2 stages across two machines.
  • SIWS + 2FA security: For public deployments, operator access is a Sign-In With Solana signature over a server nonce, plus TOTP 2FA and single-use backup codes. A device that only wants to contribute compute never signs in: it proves a wallet against the bridge and receives a token scoped to participation and nothing else.

What is running, and what it measured

Measured on our own machines

  • MoE serving today — Step-3.7-Flash, across two NVIDIA GB10 machines
  • stages the model is split into, placed across two agents
  • cosine similarity across ROCm, CUDA and a phone CPU — mixed hardware agrees
  • wallet platforms — web · desktop · iOS · Android
  • Step-3.7-Flash 198B at 6.7–6.8 tok/s, first token in 67–71 ms · 122B end to end across three machines · OpenAI + Anthropic compatible · Solana devnet

Live today, and where it’s going

A clear line between what already runs and what is planned. We don’t present the roadmap as shipped.

  • Now — Frontier-scale serving, live: Step-3.7-Flash — a 198B MoE — is placed as 2 stages across two NVIDIA GB10 machines and is serving. A 122B model ran end to end across three physical machines, one of them a phone holding part of it. Wallets on web, desktop, iOS and Android keep keys on the user's device.
  • In flight — Ring, gateway, client: The settlement gateway is ported to p4 and answering: a paid request crosses the gateway, the bridge and the GB10 ring and is billed on the tokens it used. On the serving ring, Step-3.7-Flash returns 6.7–6.8 tokens a second single stream, first token in 67–71 ms. Still in flight: the ring is going through recovery gates after an interrupted 64-request run, the desktop client is becoming a real node that supervises a p4 agent instead of only registering one, and expert-grain sharding is being carried onto p4 — a device can already claim a window and open a relay, but the shard download and the engine-side dispatch are not written yet.
  • Coming — Kimi K3, and settlement beyond devnet: Kimi K3 (2.8T MoE) verification on p4 is the next gate, and the GLM-5.2 report — measured on linkcpp, the previous engine — is still to be published. Settlement runs on Solana devnet with an off-chain service; an on-chain rewards program and mainnet are planned.
  • Coming — A permissionless global network: Current demos ran on one operator’s hardware. Opening the network so anyone, anywhere can plug in a device and earn — with no gatekeeper — is next.