Engineering notes, release reports, and protocol writing from the team building the AI-native post-quantum L1. Concrete, verifiable, honest about gaps.
For the first time, a fleet of ordinary 8GB GPUs trained an improvement into the model qbc.network serves, and it passed a full evaluation, went live, was attested on-chain, and its weights are public. Here is the milestone, and why the method points at something much larger.
A model improvement trained by a fleet of ordinary 8GB GPUs has passed a full uncapped evaluation and been promoted into the model qbc.network serves: generalisation up 0.3405 nats, MMLU flat, GSM8K up 2.50 points. What we ran, what we measured, and what we did not overclaim.
QBC's pitch is that the model's improvement is provable on-chain. We found a place where it was not: every training epoch's attestation signed a hardcoded loss constant, not the real measured value. The gap, the fix now live, and the parts we deliberately did not fake.
We hardened the DiLoCo aggregator that turns many workers' gradients into one model. New defenses make a single malicious or faulty worker unable to poison training or stall the loop. Here is exactly what changed, and what it does not yet cover.
The attested decentralized training loop crossed train step 1,000,000, finalized by validator quorum, with held-out loss down 64% across the live series. What the number counts, and what it does not.
Multi-worker decentralized AI training is live: three independent machines train one model, each signing work with a post-quantum key and earning QBC. One has no GPU at all, and it earns anyway.
The first on-chain reward for decentralized AI training is live on QBC: a model checkpoint committed, attested by validator quorum, finalized, and paid. Every step is checkable on-chain, fully automatic.
A real DiLoCo decentralized training run across two physical machines: a GPU box and a CPU box exchanging only small adapter updates over the network. Momentum accumulates across rounds and the global model beats base.
The 7B model serving QuantumAI Blockchain now trains in process on a single 12 GB consumer GPU: layer-wise LoRA, multi-worker DiLoCo, and a Sephirot MoE. Every surface beats the frozen base under a regression guard.
FedAvg forgets everything between rounds; DiLoCo syncs workers through an outer momentum buffer that remembers the descent direction across rounds. We replaced our FedAvg aggregator with the DiLoCo outer optimizer, tested and shipped in production.
Aether v7.1 end to end: the Sephirot adapter trained on the 7B serve-target, published, served in-process, and its identity root attested on-chain. One model now answers chat, produces phi, and is what the chain records. Every number and hash is verifiable.
Aether v7.1-unified collapses three models into one: it generates chat, exposes its attention for phi, and improves every Aether domain, zero regressions. One in-process model on a frozen 7B base, tracked by the chain, downloadable by you.
Decentralized AI training that pays: run the client, contribute your GPU, and every epoch that measurably improves AethersMind pays out in QBC. How the loop works, what is live today, and why the incentives hold as the network grows.