İçeriğe geç
MZGrid

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

MZGrid mainnet beta is live on Base

The escrow, provider registry and MZG token contracts are deployed on Base mainnet, the marketplace is open, and providers can start serving jobs. Here is what shipped, how payment works, and what is deliberately still limited.

JL

Jonas Lindqvist

Ürün

· 7 dk okuma

Today we are opening the MZGrid mainnet beta on Base. Customers can rent GPUs from independent providers, pay in MZG through an on-chain escrow, and get an endpoint back in a few minutes. Providers can install the agent, benchmark their cards and start accepting jobs the same day.

This post covers what shipped, how a job moves from deposit to settlement, and — just as important — what is still limited during the beta. We would rather you hear the caveats from us first.

What is live today

The beta consists of five pieces that ship together:

  • Contracts on Base mainnet. MZGToken (a fixed-supply ERC-20 with EIP-2612 permit), ProviderRegistry (provider status, the 20% lock and slashing) and MZGridEscrow (job deposits and settlement). The contracts are not upgradeable. Addresses are listed on the contracts page.
  • The coordinator. A REST API that runs the scheduler, verification and settlement. It is what the website, the SDK and the provider agent talk to.
  • The marketplace. Live offers on the marketplace, filterable by GPU model, region, VRAM and reliability, with a deploy flow that walks you through the escrow deposit.
  • The provider agent. The mzgrid-provider CLI detects your GPUs, runs the registration benchmark, registers on-chain and runs jobs in isolated containers. Installation is covered in the provider docs.
  • The TypeScript SDK. @mzgrid/sdk wraps the API and the contract calls. See SDK docs.

At launch the catalogue covers thirteen NVIDIA models, from the GeForce RTX 3090 up to the B200, each with a reference price in MZG per GPU-hour. The full table is in the GPU catalogue. In practice, the first providers are mostly the datacenter and professional operators who ran machines on the testnet — A100, H100 and L40S servers next to RTX 4090 and RTX 3090 rigs — so early supply leans towards larger, more expensive cards. We expect community providers with one or two consumer GPUs to make up most of the network over time, and the marketplace will show that mix as it changes rather than a picture we would like it to be.

The first templates are:

  • vllm-openai — an OpenAI-compatible endpoint for any Hugging Face model that fits on the card
  • llama-3.1-8b — Llama 3.1 8B Instruct on vLLM, our recommended first job
  • sdxl-comfyui — ComfyUI with SDXL and FLUX.1-schnell
  • pytorch-jupyter — PyTorch 2.5 with CUDA 12.4 and JupyterLab
  • custom-docker and batch — bring your own image

How a job is paid for

Everything on MZGrid is priced and paid in MZG. There is no second currency, no price feed and no conversion step. A job goes through four stages:

  1. Deposit. You pick an offer and a maximum duration. The escrow contract takes pricePerHour × maxDuration from your wallet before anything runs.
  2. Accept. The provider has 10 minutes to accept the job on-chain. If they don't, you can cancel and get the full deposit back.
  3. Run. The container starts, the agent reports GPU utilisation every 60 seconds, and you get your endpoint URL.
  4. Settle. When the job finishes or you stop it, the coordinator settles it on-chain. You pay for the time you actually used; the rest of the deposit comes back to you in the same transaction.

Of the amount charged, 7% goes to the protocol treasury. The remaining 93% is the provider's share: 80% of it is paid out immediately and 20% is locked in the provider registry. That lock is the provider's collateral — there is no separate staking step — and it is released only when a provider leaves the network, after a 10-day exit delay. Our escrow and billing docs walk through the arithmetic.

If a running job is never settled, you are not stuck: once the deadline plus a one-hour grace period has passed, you can call reclaimTimeout yourself and get the full deposit back without involving us.

How providers join

There is no allow-list. Anyone with a supported NVIDIA GPU, a stable connection and a Base wallet can register. Two mechanisms keep this safe for customers:

  • A benchmark gate. At registration, each card runs an FP16 matrix-multiply benchmark. The result has to land between 70% and 115% of the catalogue reference for the claimed model, or the machine is marked unverified and receives no jobs.
  • A job cap that grows with the lock. A new provider can only take jobs with a deposit of up to 25 MZG. Each completed job adds to the provider's lock, and the cap grows as 25 MZG + 4 × locked. Providers earn their way into larger jobs; nobody needs to buy their way in.

Providers set their own prices per GPU model. Offers typically sit between 0.85× and 1.20× of the reference price. The pricing guide explains how the scheduler weighs price against reliability.

What is deliberately limited in the beta

We are calling this a beta because several parts of the system are intentionally conservative or centralised today. You should know about all of them before you deposit MZG or register a machine.

Uyarı

The coordinator — scheduler, verifier and settlement oracle — is operated by MZGrid Labs. It can settle jobs, fail jobs and submit slashing transactions. The contracts bound what it can do (it can never charge more than your deposit, and slashing is capped at a provider's locked balance), but it is a trusted component. Read the risk disclosure before using the network.

  • No external audit has been completed. The contracts have a full unit, fuzz and invariant test suite and an internal review, documented in Security. That is not the same as an independent audit, and we are not claiming one.
  • Admin roles sit with a multisig. The admin multisig can grant and revoke the oracle and slasher roles. There is no token voting and no governance process; the multisig's powers are documented on the contracts page.
  • Not confidential computing. Your container runs on someone else's hardware. Isolation protects providers from jobs and jobs from each other, but a provider with physical access to a machine can in principle inspect its memory. Do not send data to a job that you would not be comfortable with a third party seeing.
  • Inference first. Templates are tuned for inference endpoints and short interactive sessions. Long multi-node training is not supported; every job runs on a single machine.
  • Two transactions to deploy. The web app currently asks you to approve MZG and then fund the escrow. The token supports EIP-2612 permits and the escrow has a createJobWithPermit function, but the web flow does not use it yet.
  • Minute-rounded settlement. In the beta, the coordinator rounds used time up to the next whole minute before settling. The contract itself bills per second; we want more production data on usage reporting before we pass exact seconds through.

We expect to lift some of these limits over the coming weeks and months. We will announce changes here and on the status page — and we will not promise dates we can't keep.

Where the network starts

During the public testnet on Base Sepolia, a few hundred machines registered and ran test workloads, which gave us most of the bug reports that shaped the agent. In the first hours after launch, 30 providers across 11 countries have registered on mainnet with roughly 170 GPUs between them. That is a small network, and we know it. The network dashboard shows the live numbers, unedited, from the coordinator's own statistics endpoint.

Getting started

The code is on GitHub. Questions, bug reports and blunt feedback are welcome on Discord and Telegram, and we post operational updates on X.

Thank you to everyone who ran machines on the testnet. The beta is the start of the work, not the end of it.

Blogdan daha fazlası

Tüm yazılar

Ağa katılın

GPU’larınızı getirin. Modellerinizi getirin.

Dakikalar içinde bir çıkarım uç noktası dağıtın ya da atıl donanımınızı MZG kazancına dönüştürün. Sağlayıcılarımız ve geliştiricilerimiz Discord’da buluşuyor.