Decentralised AI networks, compared

The decentralised-AI landscape

A cluster of projects is decentralizing pieces of the AI stack. Bittensor runs subnets that incentivize off-chain machine-learning work, largely inference, ranked by peer evaluation. Ocean Protocol builds marketplaces for data and compute. Others focus on GPU rental (DePIN) or agent frameworks. Each decentralizes something valuable, and conflating them helps nobody: a data marketplace, an inference market, and a training network are different products with different failure modes. The wider field, including newer entrants and how they position, is surveyed on AI crypto projects in 2026 and, for the infrastructure side, on the decentralised compute landscape. QBC's contribution is different in kind from all of these, which is the point of this comparison rather than a dismissal of the field. The broader thesis is on the Bitcoin-for-AI page.

What QBC does that the others do not

QBC finalises the trained model itself on-chain. The weight root is the on-chain object that validators attest, and rewards reference that finalised improvement, verified against a held-out loss by a two-thirds quorum. Most decentralised-AI networks settle a summary, a ranking, or a market transaction; QBC settles the model. That is the difference between decentralizing the economy around AI and decentralizing the artifact. The finalised checkpoints accumulate in an on-chain model registry, so the artifact has a public, tamper-evident release history rather than a mutable model card, and any party can bind a benchmark claim or an audit finding to an exact attested checkpoint. The full mechanism, from local training step to finalised epoch, is on the decentralised AI training page.

Verification: peer ranking vs. held-out finality

How a network decides whose work was good determines whether it can be gamed. Incentive networks often use peer evaluation, which is powerful but vulnerable to collusion at the ranking layer, since the evaluators are also economic participants with positions to defend. QBC verifies against a held-out set the trainer never saw, and finalises only improvements a validator quorum confirms; a regressing or poisoning epoch is rolled back and earns nothing. Held-out verification has its own limits, and QBC states them: it measures the model, not the intent, so it filters harmful updates by their effect rather than their motive. The general problem, how a network can check that training actually happened and actually helped, is treated on validating ML training; the consensus design that anchors QBC's answer is covered under Proof of Thought.

Post-quantum security as a differentiator

Most decentralised-AI networks inherit the classical cryptography of the chains they build on, leaving contribution and reward signatures exposed to a future quantum computer. QBC is post-quantum from genesis (NIST FIPS 204): every contribution is Dilithium-signed, node-to-node transport is wrapped in ML-KEM-768, and every reward flows to a post-quantum identity. For a network meant to last, this is a structural advantage rather than a feature checkbox, because migrating a live incentive network to new cryptography after the fact means re-keying every participant and revalidating every assumption downstream. The signature scheme itself, and its practical costs on-chain, is unpacked on the Dilithium signatures page; the full model, from keys to transport to consensus, is on the post-quantum blockchain page.

Verifiability of the served model

In many networks you cannot prove which model produced a given output. On QBC the served model hashes to an on-chain manifest root matching its public release, and its reasoning is committed per block. That makes the whole system auditable, not just the incentive layer (see verifiable AI). It is a different design goal: not only reward decentralization, but provenance. QBC also publishes real benchmark deltas rather than vibes: the engineering log AethersMind v7: real benchmarks reports GSM8K up roughly ten points against the frozen base while MMLU stayed flat, an honestly mixed result stated as measured, against a checkpoint the chain attests. A comparison shopper should ask every network in this category the same question: which exact model earned these numbers, and can you prove it?

The training method: DiLoCo, not subnet scoring

Under the incentive layer sits a more basic difference: what kind of distributed learning the network actually runs. Bittensor subnets largely reward off-chain inference and ranking; QBC runs a genuine distributed training loop based on DiLoCo, in which workers train locally for many steps and exchange kilobyte-scale sparse, signed pseudo-gradients as outer updates. That makes wide-area training feasible on ordinary links and makes every contribution attributable to a post-quantum identity. The method and its trade-offs against classical federated averaging are compared on DiLoCo vs federated learning, and the implementation history is public rather than claimed: multi-worker decentralised training records the step from one worker to several concurrent workers feeding a single shared model on the live network. The distinction to hold onto: an incentive market can exist without any shared model improving, while QBC's loop is pointless unless the shared model measurably learns, which keeps the incentives and the artifact aligned.

What is live today, with numbers

An honest comparison should state scale. QBC's chain currently runs four validators across three physical machines; the thousand-node network is roadmap, and QBC labels it as such. Against that honest denominator, the live numerators are substantial: the shared model has crossed one million training steps (train step one million), held-out loss is down more than 75% since the loop began, and training rewards have been paid end to end, from submitted pseudo-gradients through two-thirds quorum attestation to automatic on-chain payouts, documented in training rewards live: the first paid epoch. The served model is a frozen Qwen2.5-7B base with a ten-expert Sephirot adapter, attested on-chain at block 1,097,990 with open weights on HuggingFace. Larger networks in this comparison have more nodes and bigger markets; QBC's claim is not size but verifiability per unit of scale. A similar honesty-first framing is applied to compute-verification networks on the Gensyn comparison.

How to choose, honestly

If you want decentralised inference incentives, or a data or GPU marketplace, mature options exist and QBC is not trying to replace them. If you want a decentralised model that is trained, verified, and finalised on-chain, with post-quantum security and verifiable provenance, that is QBC's specific contribution. It is complementary to the rest of the stack: you could source compute from DePIN (DePIN for AI) and still finalise the model on QBC. The decision is really about what you need to be able to prove afterward: market participants need settlement, while builders of accountable AI systems need provenance, and the two needs favor different networks. If participating rather than comparing is your goal, the practical path is on joining the network. Read the honest current state, including what is live and what is roadmap, in the QBC whitepaper.