İçeriğe geç
MZGrid

Bu sayfa yalnızca İngilizce olarak mevcuttur. Bağlayıcı olan İngilizce sürümdür.

Core concepts

Offers, SKUs, templates, escrow, settlement, provider lock, reputation, job caps, regions and provider tiers explained.

This page defines the vocabulary used throughout the docs, the API and the contracts.

SKU

A SKU is the stable identifier of a GPU model, such as h100-sxm, a100-80 or rtx4090. SKUs are used in the API, in provider pricing and in the scheduler. Each SKU has a fixed VRAM size, architecture, an FP16 reference throughput used for benchmark verification, and a reference price in MZG per GPU-hour. See the GPU catalogue.

Offer

An offer is a provider's listing for one machine: the SKU, how many GPUs the machine has (1, 2, 4 or 8), how many are available right now, CPU, RAM, disk and bandwidth, the region, and the provider's price per GPU-hour in MZG. Providers set their own prices; live offers typically sit between 0.85× and 1.20× the reference price. A job rents one GPU from an offer.

Template

A template is a pre-configured workload: a container image, the ports it exposes, a minimum VRAM requirement and a list of recommended SKUs. Templates include vLLM endpoints for Llama 3.1 8B, Llama 3.3 70B and Qwen2.5 32B, a generic vLLM endpoint for any Hugging Face model, ComfyUI, Whisper, JupyterLab, and the custom-docker and batch templates for your own images. See Templates.

Escrow

Every job is funded before it starts. The customer deposits pricePerHour × maxDuration into the MZGridEscrow contract. Neither the provider nor MZGrid can withdraw the deposit directly: it can only leave the escrow through settlement, cancellation, failure or timeout reclaim, each governed by the rules in MZGridEscrow.

Settlement

When a job ends, the coordinator — holding the escrow's ORACLE_ROLE — calls completeJob with the number of seconds used. The contract computes:

Metin
charged       = min(deposit, pricePerHour × usedSeconds / 3600)
fee           = charged × 7%            -> treasury
providerShare = charged − fee           (93% of charged)
locked        = providerShare × 20%     -> ProviderRegistry (18.6% of charged)
liquid        = providerShare − locked  -> provider wallet  (74.4% of charged)
refund        = deposit − charged       -> customer

Lock

The lock is the provider's collateral. MZGrid has no staking step: instead, 20% of every provider payout accumulates in the ProviderRegistry. Locked MZG can only be withdrawn after the provider requests exit and waits 10 days with no active jobs. During that time it can still be slashed. See Lock & exit.

Job cap

A new provider with nothing locked cannot take arbitrarily large jobs. The registry enforces a per-job maximum deposit:

Metin
jobCap = 25 MZG + 4 × lockedBalance

A provider with no lock can accept jobs with deposits up to 25 MZG; one with 1,000 MZG locked can accept jobs up to 4,025 MZG. The cap grows automatically as the provider earns, so collateral always scales with the size of the jobs entrusted to them.

Slashing

If a provider abandons an accepted job, goes offline for more than 10 minutes during a running job, or fails verification, part or all of their locked MZG is slashed. Half of the slashed amount goes to the affected customer, half to the treasury. See Slashing.

Reputation

Each provider has a reliability score from 0 to 100, computed as 40% completion rate over 90 days, 30% uptime over 30 days, 20% benchmark consistency and 10% tenure (capped at 180 days). It is shown on every offer and used by the scheduler. See Reputation.

Regions

Offers are tagged with a region code such as us-east, eu-central, kr-central or sg-central, plus the country and city. Choose a region close to your users to reduce latency; the scheduler's proximity term does this automatically when you use autoSelect. The live region list is available from GET /v1/network/regions.

Tiers

Providers are labelled with one of three tiers:

TierTypical profile
datacenterHosted in professional facilities with redundant power and networking, usually Hopper/Blackwell or A100 hardware. Prices tend toward the upper end of the live range.
proDedicated operators running small fleets of workstation or server GPUs with business-grade connectivity.
communityIndividuals contributing consumer GPUs such as the RTX 4090 or 3090. Often the most affordable option for 8B–32B inference and experimentation.

Tier is descriptive. Every tier is subject to the same benchmark verification, job caps, lock and slashing rules.

Job lifecycle

StatusMeaning
awaiting_fundingCreated via the API; escrow not yet funded.
fundedDeposit is in escrow; waiting for the provider to accept.
startingProvider accepted on-chain and is pulling and starting the container.
runningContainer is up; the endpoint (if any) is live and usage is being metered.
completedSettled with completeJob; unused deposit refunded.
failedSettled with failJob; full refund.
cancelledNot accepted within 10 minutes and cancelled by the customer; full refund.
reclaimedNot settled by deadline + 1 h; customer reclaimed the full deposit.