vearlBitcoin mines hashes. Vearl mines AI.
Vearl keeps the security of proof of work, but miners earn it by running AI models for real clients. Every answer is proven correct before anyone gets paid.
Closed testnet today · public testnet planned October – November 2026
Bitcoin's network draws as much power as a mid‑sized country, to compute hashes that nobody will ever read. Vearl points that power at work people actually need.
Same hardware, same energy. A result at the end.
In a classic proof of work, miners compute billions of hashes whose only purpose is to be hard to find. Vearl keeps the security model and changes what gets computed: every block is mined from a matrix product, and that product can belong to a paying client.
Integer arithmetic is deliberate: every node reproduces a result bit for bit, which is what makes cheap verification and one-operation fraud proofs possible.
Running on the testnet today.
Every figure below was measured on the testnet, on a single NVIDIA A10 GPU, with each result checked before payment. Nothing here is a projection.
| Workload | Model | Size | Measured |
|---|---|---|---|
| AText generation | TinyStories-33M language model, 4 layers, 50,257-word vocabulary | 307 ops · 2 jobs per word | Writes stories word by word: 8 stories in parallel, ≈ 18 s per word, 92.9% same next word as the original model |
| BPhoto recognition | DeiT-Tiny vision transformer, 224 × 224 photos, 1,000 ImageNet classes | 386 ops · 2 chained jobs | ≈ 1.7 s per photo, 70.6% top-1 on photos never seen in training (original model: 78.2%) |
| CSemantic search embeddings | BGE-base, 12 layers, 110 million parameters | 492 ops · 3 chained jobs | ≈ 3.3 s per text, 96.7% of the original model's score on STS-B (0.836 vs 0.864) |
| DFast sentence embeddings | all-MiniLM-L6-v2, 6 layers, 12 attention heads | 195 ops · 61 GEMMs | ≈ 1.0 s per text, cosine similarity 0.92 with the original model |
| ESmall-image classification | ResNet-20 on CIFAR-10 | 76 ops · 22 GEMMs | 88.6% top-1 accuracy, 32 images in 41 s |
| FMatrix multiplication | Tiled int8 GEMM | 8192 × 8192 × 8192 | 1.1 trillion operations, verified end to end in about 11 s |
| GHandwritten digits | MLP and CNN on MNIST | 10,000 test images | 98.1% top-1 accuracy in int8 |
| HBulk embeddings | Distilled text encoder | 20,480 texts | Outputs bit-identical to an independent reference engine |
| IYour own model | PyTorch or ONNX import | 240 ops per job, chainable | Convolutions, attention, LayerNorm, softmax, GELU, residual blocks |
Models are quantized to 8-bit integers so that every node computes exactly the same bits. Accuracy figures compare the integer model with its original floating-point version on held-out data. These models are a first generation: they will be considerably improved, with larger and more accurate models, faster jobs and longer texts.
Nobody trusts the miner. Anyone can check.
Redoing a computation costs as much as the computation itself. Freivalds' test checks a matrix product with a few matrix-vector products instead, and a single wrong entry is enough to prove fraud.
The client submits a job
The input and the Merkle roots of the model weights go on chain, together 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, exact recomputation for everything else. Seconds, not the cost of the original job.
Payment or penalty
All correct: the escrow is released. One wrong operation: a fraud proof covering that operation alone is enough, and the miner loses its bond.
Live illustration on 6 × 6 matrices. One job in three, the miner alters a single cell.
Standard cryptography, independently cross‑checked.
No homemade primitives. Each component follows a public standard or a well-known reference, and is tested against a second implementation written separately from the specification.
Post-quantum accounts
SLH-DSA-SHAKE-128s (NIST FIPS 205) alongside Ed25519, checked against an independent implementation of the standard.
ASERT (aserti3-2d)
Continuous difficulty adjustment, validated against the 29 official reference vectors.
Atomic reorgs, optional MESS
State snapshots with exact rollback, plus an optional brake on deep reorganizations using ECIP‑1100 parameters.
Header, tile and balance proofs
Verify the chain and an account balance without downloading blocks or job data.
Noise XX over X25519
Encrypted and authenticated, but still classical. A post-quantum handshake is not implemented yet.
External security audit
17 independent checks and differential fuzzers run on every change. No third-party audit has been done yet.
Where the project stands.
Vearl is in development and runs as a closed testnet. Testnet tokens have no value. We publish what works, with measurements, and what does not work yet.
automated tests, all passing
independent checks in the internal audit suite
conformance blocks shared between implementations
external audits so far
- Done
Closed testnet
Chain, compute market, verified inference and embeddings on a GPU server.
- October – November 2026
Public testnet
Multiple machines, external miners and a block explorer. Miners who help secure it are planned to receive Vearl tokens on mainnet.
- Next
External audit
Third-party review of the cryptography, the consensus rules and the fraud proofs.
- Research
Considerably better models
Larger and more accurate models: per-channel quantization, several sequences per job, an on-chain model registry.
Straight answers.
The questions people ask first, answered without marketing.
Is Vearl live?
Vearl runs as a closed testnet on a GPU server. The next milestone is a public testnet that external miners and developers can join, planned for October to November 2026. 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 still have no value.
Who is behind Vearl?
Vearl is created and led by Clovis Anicet, a publicly identified (doxxed) founder: the project is not anonymous, and you can write to contact [at] vearl.org. A name alone proves nothing, so the site also publishes the measurements, the limits and what is not solved yet, which anyone can check.
Do testnet tokens have any value?
No. Testnet tokens exist only for testing. They have no monetary value and are not for sale.
What hardware do I need to mine?
GPU mining needs an NVIDIA card with compute capability 8.0 or newer: Ampere and later, for example the RTX 30 and 40 series, A10, A100 or H100. Older cards are refused by the node. CPU mining works for testing.
How is this different from Bitcoin?
The security principle is the same: producing a block costs real computation. What changes is the computation itself. Bitcoin miners compute SHA-256 hashes; Vearl miners compute integer matrix products, which can carry useful work for clients.
Can it run large language models?
Small ones, not large ones. The testnet runs a 33-million-parameter language model (TinyStories) that writes short stories word by word, every step verified before payment, and a 110-million-parameter text encoder. A job holds up to 240 operations on 1024 × 1024 tensors and jobs chain together, which reaches models of roughly a hundred million parameters. Models with billions of parameters are out of reach today, and text is limited to 64 tokens.
Has the code been audited?
Not by a third party yet. Internally, 17 independent checks and differential fuzzers compare the Rust node with separate Python implementations written from the specification, on every change.
Your GPU could be doing something useful.
The public testnet will open to miners and developers first. Miners who help secure the network are planned to receive Vearl tokens on mainnet as a thank-you. Everything it does is documented above, with the numbers behind it.
Questions, press or partnerships: contact [at] vearl.org