AI Sharing¶
The AI Sharing page (Advanced mode) prepares this PC as an AI-sharing provider. It is entirely opt-in.
Connect to a local runtime¶
AI Sharing connects to MeshLLM — or any OpenAI-compatible runtime — that is already running on
http://127.0.0.1:PORT/v1. Only literal loopback is accepted, so the wallet never reaches out to a remote AI
endpoint on your behalf.
You set an opt-in provider policy: inference and hosting consent, GPU/CPU/power/temperature/bandwidth limits, the UTC hours you serve, a minimum price and a payout address. The policy can be exported for the MeshLLM supervisor.
"Ask the model" sends a test request and shows a local usage observation. That observation is unverified and not payable: AI-sharing payouts need authenticated receipts and settlement that Core does not provide yet, and the wallet never mints or claims rewards from self-reported usage.
Run the local Bonsai model¶
The page can download and run Ternary Bonsai 2 2B (Prism ML, Apache-2.0) on this PC, in five visible steps: Download → Verify → Install → Ready → Sharing active.
The rules that keep this safe:
- Consent first. Download and hosting are separate tick boxes, both off by default and saved per PC. Nothing is downloaded or started without them.
- Pinned sources. A signed bundle (
pqwerai-local-model-bundle/1) pins the exact model revision and every file's size and SHA-256, and pins the llama.cpp runtime build by SHA-256. The default variant isQ2_0(547 MB). - Download rules. Downloads use HTTPS only, from Hugging Face or GitHub hosts only (including redirects). They resume after an interruption and can be paused. A file with a wrong hash is deleted.
- Registry check. A version is installed and started only while the on-chain ModelRegistry lists it as
approved and not revoked (
servingAllowed && !disabled), shows the bundle's signer as the model publisher, and matches the manifest digest, weights and licence. A revoked version is refused at install and at every start. - Runtime limits.
llama-serverlistens on127.0.0.1only, with the Web UI off and the model served under an alias (no local paths exposed). It runs at below-normal priority inside a Job Object with a memory limit, is killed with the wallet, and restarts automatically at most 3 times in 10 minutes. A real chat completion is checked before the page shows Ready. - Execution. Nothing downloaded is executed except the verified
llama-server.exe— no scripts run and no private keys are read. - Images. The vision projector (
mmproj-Q8_0, 365 MB) is downloaded only when you enable images. - GPU. Vulkan GPU offload is used only when you allow it and the card has enough memory.
- Updates and rollback. Install signed update… accepts a newer approved bundle of the same model and reuses already-verified weights; the previous version is kept, and Roll back switches back while the registry still allows it to serve.
- Hardware fallback. Below the manifest minimum (x64, AVX2, 4 GB RAM, 2 GB free disk), local hosting is disabled and the page offers an approved provider instead, paid per job — see Paid Ask and AI escrow.
- Storage. Models and runtimes live in
%LOCALAPPDATA%\PQWERAI\Wallet\ai\local, outside the roaming profile.
Priority over mining¶
If your own node is mining and AI Sharing is active, you can let AI sharing come first: the node and its helper drop to Idle CPU priority while sharing is active (CR-0053, an option on this page). See mining.
Payments are not live yet
The AI-sharing revenue split exists as a policy, but authenticated receipts and settlement are not provided by Core yet, so self-reported usage is never paid. Paid work today goes through the escrow flow on Paid Ask and AI escrow.
Sources at 7a33c77104: apps/pqwerai-wallet-windows/README.md
Found a problem? Report a documentation issue · Build testnet-rc1+7a33c77104