[ANN] Vearl | Proof of Useful Work: miners run verified AI models Public testnet

https://vearl.org/og-image.png?v=3

Vearl
Proof of Useful Work · Verified Compute
Bitcoin mines hashes. Vearl mines AI.

Website: https://vearl.org
Whitepaper: https://vearl.org/whitepaper
Discord: https://discord.gg/jya8meKnvM
Contact: contact@vearl.org
Team: public (doxxed), Clovis Anicet (founder) and Idan Sugarman
Public testnet: planned for 23 October 2026

What is Vearl?

Vearl is a proof-of-work blockchain where the work that secures the chain can also be a paying client's computation. Instead of hashing for hashing's sake, miners multiply 8-bit integer matrices on GPUs. Every 16 x 16 tile of a noised product is a lottery ticket, and a ticket below the difficulty target is a block. The same matrix products can belong to a client's AI job, so the energy that secures the network also produces a result someone needs.

Results are exact integers, so every node computes identical bits. That is what makes cheap verification and one-operation fraud proofs possible.

A 30-second film is on the home page of the website.

How a job works

The client submits a job. The input and the Merkle roots of the model weights go on chain, with an escrow that will pay the miner.
A miner runs it on a GPU. It executes every operation, publishes a commitment to each result, and mines blocks from those same matrix products.
The client verifies. Freivalds' test for every matrix product (worst-case error 2^-24 with 24 rounds), exact recomputation for everything else. It costs seconds, not the price of the original job.
Payment or penalty. All correct: the escrow is released. One wrong operation: a fraud proof of about a kilobyte covers that operation alone, and the miner loses its bond.

What already runs (measured on the closed testnet, on one NVIDIA A10)

Text generation: TinyStories-33M (33M parameters), every word verified. 8 stories in parallel, about 18 s per word, 92.9% same next word as the original model.
Photo recognition: DeiT-Tiny, about 1.7 s per photo, 70.6% top-1 on unseen photos (original float model: 78.2%).
Text embeddings: BGE-base (110M parameters), about 3.3 s per text, 96.7% of the original model's score on STS-B. all-MiniLM-L6-v2 at about 1.0 s per text.
Also: ResNet-20 (88.6% on CIFAR-10), MNIST (98.1%), and an 8192 x 8192 x 8192 tiled matrix product (1.1 trillion operations, verified end to end in about 11 s).

Models are quantized to 8-bit integers so that every node computes the same bits, which costs some accuracy (figures above compare the integer model with its float original). A job holds up to 240 operations on 1024 x 1024 tensors and jobs chain together, which reaches models of roughly a hundred million parameters. These models are a first generation: they will be considerably improved.

Chain and cryptography

30 s target block time, ASERT (aserti3-2d) difficulty adjustment, validated against the official reference vectors.
Post-quantum accounts: SLH-DSA-SHAKE-128s (NIST FIPS 205) alongside Ed25519.
Atomic reorgs with state rollback, optional MESS brake on deep reorganizations (ECIP-1100 parameters).
Light clients: header, tile and balance proofs.
Written in Rust and CUDA. 171 automated tests, 17 internal audit checks, 483 conformance blocks shared between independent implementations.

Team

The team is public (doxxed): Clovis Anicet (founder) and Idan Sugarman. You can reach us at contact@vearl.org or on Discord.

Roadmap

Done: closed testnet (chain, compute market, verified inference and embeddings, text generation).
23 October 2026: public testnet, with several machines, external miners and a block explorer.
Next: external security audit.
Research: considerably better models (larger, more accurate, faster), per-channel quantization, an on-chain model registry.

Miners who help secure the public testnet are planned to receive Vearl tokens on mainnet as a thank-you. This is a plan, not a guarantee: the rules and amounts will be announced, and there is no mainnet date yet. Testnet tokens themselves have no value and are not for sale.

Honest limitations

No external audit yet. Several security arguments are empirical or simulated, not proven.
Models are small and lossy (no billion-parameter models), text is limited to 64 tokens.
The P2P channel (Noise XX over X25519) is classical, not post-quantum yet.
So far everything has been measured on a single machine, and the 3D rendering shown on the website is illustrative (the testnet kernel draws simpler scenes).

Mining requirements

NVIDIA GPU with compute capability 8.0 or newer (Ampere and later: RTX 30/40 series, A10, A100, H100). CPU mining works for testing only.

Community

Join the Vearl Discord to follow the news, ask questions and get ready for the public testnet: https://discord.gg/jya8meKnvM

Questions and criticism are welcome, and we will answer them here and on Discord. Everything above is documented with numbers in the whitepaper.

This is testnet software. Testnet tokens have no value and are not for sale. The code has not been audited by a third party. Nothing in this post is investment advice.