Engineering notes, release reports, and protocol writing from the team building the AI-native post-quantum L1. Concrete, verifiable, honest about gaps.
Native QBC now moves between the quantum-secure L1 and the QVM on chain 3303. Every QBC on the QVM is backed by exactly one QBC in L1 escrow, shows up in MetaMask like any token, and goes back to L1 any time. Try it at qbc.network/bridge.
A long look at the QV desktop wallet: a post-quantum cold-storage wallet that holds Bitcoin, Lightning and QBC behind one recovery phrase. This is a private build, not a public release, and this post explains what is in it and how we harden it before anyone trusts it with value.
Three tiers of on-chain privacy, all demonstrated live today with the transactions to prove it: amount-hiding payments, post-quantum amount-hiding with real value, and a post-quantum unlinkable spend that reveals only a nullifier. The walkthrough, the evidence, and how to run it.
We designed and built the stack to serve a 2.8T-parameter open MoE on ordinary 12GB consumer GPUs, paid per token on-chain, owned by no one and served by everyone. The architecture, the honest memory maths, and what is built versus what still needs a fleet.
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.
Most AI safety is a promise. On QBC it is a transaction. An end-to-end walk through the safety path: the Layer 1 root of trust, the Gevurah and Malkuth contracts on the QVM that make a veto enforceable, and how one chat turn leaves a permanent, auditable record.
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.
Chat on the live 7B went from sixty seconds to five: three stacked causes found and fixed, decode up from 1.6 to 30 tokens per second. KV quantization stays ON and a public chat answers in five to seven seconds.
A 200x performance fix on a live 5-million-vector fabric: every search had degraded to 29 seconds from one false assumption in index accounting. Search is now 145 milliseconds; validation sweeps went from 7 minutes to 3 seconds.
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.
BEEFY finality proofs are live on QBC: validators are signing portable commitments just behind GRANDPA on every block. Hours after shipping armed and dormant, the proofs QBC was built to hand out are flowing.
QBC ships BEEFY and MMR: an additive finality proof layer that lets a phone or another chain verify QBC finality from one signed commitment plus a proof. The groundwork for trustless bridges and thousand-voter finality.
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.
TurboQuant KV cache quantization arrives in Aether: a 7.64x smaller cache, a fused decode path, and bandwidth-optimal CUDA kernels. An honest look at what it buys today versus what it sets up for the frontier phase.
Native zero-knowledge private payments are live on QBC: amount-hiding rebuilt on Plonky3 so hiding is a guarantee of the proving system, not an argument. Reviewed, shipped value-bearing, with the on-chain record.
You can move QBC privately today: shield into a private note, transfer with the amount hidden, and cash out, all non-custodial, proven in your browser. Every protecting property is post-quantum and open for peer review.
Post-quantum private transfers live on QBC: the no-inflation proof is now a hash-based zero-knowledge STARK generated in your browser. Amounts stay hidden and soundness holds against a quantum computer.
Post-quantum Dilithium signatures and confidential transactions are live on QBC: quantum-resistant spends, and transfer amounts hidden behind commitments. Here is exactly how both work, end to end.
Confidential transactions on QBC are now permissionless end to end: shield QBC into a private note, transact with amounts hidden, and unshield back out. All non-custodial, all authorized by post-quantum signatures.
A hardening pass across consensus and the knowledge fabric: a 200x vector search fix that was slowing live chat, plus a self-healing index. Also inside: a reward drain closed in on-chain serving economics and a test that failed a healthy chain.
How warp sync lets a brand new node join the live QBC blockchain in three minutes, verifying a finality proof instead of copying a database. Proven green end to end against the live network, with one command.
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.
Autonomous GRANDPA equivocation slashing, end to end: detection, verified reports, stake attribution, and deferred slashing, wired at the runtime seam. Two stub methods had silently dropped every detected attack. This is how we reconnected the wire.
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.
V8 makes the ten Sephirot real experts: ten routed experts plus a shared one, top-k routing, and DeepSeek-V3 aux-loss-free balancing on a frozen base. The architecture V6 got wrong and V7 approximated is now built correctly, and tested.
How a silent one line bug froze VQE quantum mining for sixteen days, and the staged cutover that took single leader consensus live on production. Every block on QuantumAI Blockchain carries real VQE energy again.
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.
A full readiness review with receipts: cross-node AI knowledge search proven live, and consensus that keeps finalizing under validator churn. Two hard claims moved from built to proven. Here is the evidence.
Every few days our on-chain AI would quietly seize up: alive, still logging, answering nothing. The cause was the memory allocator hoarding freed pages. Here is the hunt, the heap-profiler proof, and the durable jemalloc fix.
Every ~500 blocks, Aether fingerprints its own cognitive state, personality traits, phi, and training loss, to an append-only on-chain ledger. How the soul ledger works, and the quiet bug that meant it had never recorded a single entry until today.
Every one of Aether's 1.07M knowledge vectors now carries a content-derived identity, and the chain anchors all ten domain roots of the fabric. Spec-142, four validators, one clean slate, the first brick of a mind no single machine has to hold.
Aether's distributed knowledge fabric answered its first cross-node search: content held only on the other machine, carried over libp2p, deduped exactly. 14/14 harness checks passed on content-addressed ids. What was built, proven, and still missing.
QuantumAI Blockchain makes AI training a public, cryptographically verified on-chain object: proven, not promised, and auditable by anyone. The leading AI labs warn powerful AI could become hard to control; we answer with verification.
The residual GRANDPA finality stall is root-caused and fixed in the node's chain-selection layer, deployed across all three validators on chain 3303. Runtime v141 ships behaviour-identical cleanup. The fix is live; the proof is still pending.
AethersMind progress report: the Knowledge Fabric crosses 1.49M vectors and every training epoch finalises on chain 3303 with a verifiable held-out loss. The v7.0 model holds general capability while sharpening domain depth.
Substrate runtime v140 is live on chain 3303 and permanently fixes the RecentProofHashes count desync behind the multi-hour finality halt at block 343742. Migration verified on live state: count 999 to 0, invariant holding.
AethersMind v7.0 rebuilds on Qwen2.5-7B with a QLoRA fine-tune: MMLU held at 69.9, GSM8K up 3.6, and Aether-domain perplexity cut 44%, all reproducible. Every number comes straight from the public model card and lm-evaluation-harness.
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.
Substrate runtime v121 is live on chain 3303: four validators bonded and elected via qbcStaking, with AuraValidatorSet reading the live staking set. Aether-Mind now runs on the RTX 3080 Ti at 10-second warm-up and 504 ms chat latency.
Every AethersMind training epoch finalises on chain 3303 with contributor set, gradient hash, model SHA-256, and a verifiable held-out loss. v6.2 is pinned at the permissionless AetherEpoch contract.
Substrate runtime v120 is live on chain 3303: permissionless staking, election, offences, and QVM-committee pallets, anyone can bond and validate. The upgrade finalised in a single block on mainnet with no missed slots.