Network

Put your GPU to verified work.

Mantissa nodes serve LLM inference that the network can prove was done honestly. At network launch, valid settled work is designed to earn $XIP; the closed alpha uses simulated credits only. Supply is admitted in step with demand so operator economics reflect real workload rather than hardware presence.

Closed alpha in progress. The public node client is not yet available. Join the waitlist and we will contact you as onboarding opens.

Node Client

Downloads

The Windows alpha pairs a native desktop operator app with the Rust node daemon and a signed inference payload. Public builds ship only after clean-install lifecycle, diagnostics, uninstall, and friend-cohort acceptance gates pass.

Linux x64

Planned for the public testnet wave, after the Windows alpha validates the onboarding and conformance flow.

Download Planned

Get notified

One email when downloads go live and onboarding opens. Nothing else.

Join the Waitlist

Operating a Node

How earning works.

At launch, Mantissa work rewards are issued only against valid settled jobs. There is no presence-based emissions faucet and no reward for idle hardware; the closed alpha uses simulated credits.

Prove your hardware

On first run, the conformance self-test checks that your GPU and driver stack reproduce the Exact Inference Profile bit-for-bit against golden vectors. Pass, and you are eligible to serve.

Bond stake, accept jobs

Operators bond stake to accept work; in-flight throughput is capped in proportion to stake, so exposure can never outrun collateral.

Serve, sign, settle

Each completed job emits a signed receipt. Receipts batch into epoch rollups; after the dispute window, valid settled work is eligible for $XIP work rewards once the token is live.

Stay honest, stay routed

Sampled audits re-execute your claimed work. Clean history builds reputation, which buys routing priority within safe bounds. Referee-confirmed dishonesty enters the full-stake penalty path.

Flaky hardware is not treated as dishonest. Consumer overclocks and non-ECC memory occasionally corrupt a bit. A single isolated mismatch rejects the job and flags it, with no penalty. Only a pattern of mismatches is treated as misbehavior. Likewise, failing the onboarding conformance test simply means the card is refused admission. That is a finding, not a fault.

Requirements

What the alpha targets.

Final requirements ship with the public build; the profile is certified per hardware generation as evidence accumulates. The alpha targets:

Component Alpha target Notes
Operating system Windows 10/11 x64 Linux planned with the public testnet
GPU NVIDIA or AMD with Vulkan support Both vendors serve through the same Vulkan backend; each card must pass the conformance self-test
VRAM 12–24 GB recommended Sized for whole-model serving of 8B–32B-class quantized models; smaller cards can qualify for smaller models
Network Reliable broadband Only the peer-to-peer port is exposed; everything else stays local
Stake Required to serve Amounts published with the token economics review

Network Telemetry

Numbers go public with the testnet.

Live utilization, verified-work volume, audit rates, and the supply-admission queue will be published here, including the share of demand that is network-funded bootstrap workload, openly labeled. Our numbers cannot lie; that is the product.

TBD active conforming nodes
TBD receipts settled per epoch
TBD sampled audit pass rate
TBD network utilization

Why a Waitlist

Supply grows with demand, by design.

DePIN networks die by letting supply grow unboundedly against no demand: earnings collapse, operators leave, the network hollows out. Mantissa runs deliberate supply admission control so operator onboarding is quota-gated and supply is admitted only as demand grows.

Utilization stays high by construction, operator earnings stay visible and real, and the waitlist itself becomes an honest growth signal.

Joining the waitlist

Email us with your GPU model, VRAM, and operating system. Alpha invitations go out in small waves as onboarding capacity and demand allow.

  • No payment, no stake, no commitment to join the waitlist
  • Every wave passes the same conformance gate; hardware qualifies on evidence, not on order in the queue
  • Your card also contributes a cross-vendor conformance datapoint the determinism literature lacks
Join the Waitlist