Business judgment on AI products
Ordinary computers join a swarm to run a huge model: you bring disk, others bring compute, and the weights never leave the machine.
01
Start inside the user's day · Public facts + workflow reasoning · 2026-08-28
Ordinary computers join a swarm to run a huge model: you bring disk, others bring compute, and the weights never leave the machine.
Public materials do not yet show how users complete this job today or what they replace.
The product targets friction in this job, but public user evidence does not yet show the cost, frequency, or consequence of leaving it unsolved.
The trend is that serving giant models may leave rented rooms for peer sharing. Don't sell a cloud-killer story. Start with lab LANs where latency is low. A standalone trend is still unclear; incentives and liability on the public net are unsolved.
It promises a simpler way to complete this job: Ordinary computers join a swarm to run a huge model: you bring disk, others bring compute, and the weights never leave the machine. The exact adoption motive and repeat use are not yet verified.
① Whether any independent third party (not the author) reproduces the speed numbers at; real internet latency; ② Whether any contributor incentive appears — P2P inference without incentives has never; run in production; ③ Whether stars and community engagement pick up — if they do not, even the deve…
Keep watching. It promises a simpler way to complete this job: Ordinary computers join a swarm to run a huge model: you bring disk, others bring compute, and the weights never leave the machine. The exact adoption motive and repeat use are not yet verified.
if you hold a compute-heavy task that is verify-able and read-heavy/write-light, outsourcing computation to a network is worth a serious look — cheaper than renting GPUs and cheap to verify. If your task is generative and subjective, cross P2P outsourcing off the list; anti-cheat costs will eat the gains.
None. Pure open-source experiment. No hosted service, no pricing page, no revenue path. ① Whether any independent third party (not the author) reproduces the speed numbers at; real internet latency; ② Whether any contributor incentive appears — P2P inference without incentives has never; run in production; ③ Whether stars and community engagement pick up — if they do not, even the deve…
Ordinary computers join a swarm to run a huge model: you bring disk, others bring compute, and the weights never leave the machine.
It promises a simpler way to complete this job: Ordinary computers join a swarm to run a huge model: you bring disk, others bring compute, and the weights never leave the machine. The exact adoption motive and repeat use are not yet verified.
An English validation note will follow from the public evidence.
The product claims to help users complete: “Ordinary computers join a swarm to run a huge model: you bring disk, others bring compute, and the w”. User evidence has not yet verified pain intensity or the cost of doing without it.
The assessment is recorded; an English explanation is pending.
The assessment is recorded; an English explanation is pending.
The assessment is recorded; an English explanation is pending.
02
Market comparison
The Chinese–English market comparison is not complete yet. A conclusion follows only after its coverage and verifiable evidence are recorded.
03
The call and next move come first; the full read retains the evidence and counterevidence.
A peer-to-peer swarm of ordinary computers running very large MoE models — your machine donates disk, another machine donates compute. Napster for LLMs.
GitHub user JustVugg (HN handle vforno), a solo developer. Apache-2.0, written in pure C. Repository created 2026-08-04; 80 commits in ten days; 47 stars / 2 forks.
It is not a fresh start: the author first built Colibrì, a single-machine project asking whether one ordinary computer can run a huge LLM (using the same MoE sparse-activation trick). Lumabri moves that question from one machine to many. The first HN post ("Show HN: Lumabri – What if LLMs worked like Napster?") drew 8 points and 10 comments. In the post the author says plainly: no datacenter, no GPU cluster, built with whatever hardware he owns.
Read: a classic solo extreme-engineering project, not a business. The real signal is the two-step sequence — first prove a MoE can be run frugally on one machine, then ask whether many ordinary machines can become one big machine. The question order is worth more than the code.
liblumibri.so (an LD_PRELOAD shim that keeps a sparse local mirror and fails over
when a peer drops)expert_node holds the weights for a few MoE
experts; other nodes send activations over the wire, the remote node computes and
sends results back. Weights never leave their home machineLUMABRI_TOKEN; LUMABRI_VERIFY spot-checks
expert calls on a second replica and demands byte-identical outputThe core claim is byte identity: the README says remote execution has been verified to produce byte-identical output to local execution on OLMoE, GLM, Inkling, Kimi K3, and DeepSeek V4.
Running a hundred-billion-parameter MoE model previously meant exactly one path: rent or buy a GPU cluster big enough. Costed by the month, out of reach for most people.
Lumabri argues for two replacements:
What it does not replace should be stated plainly: P2P distributed inference is not a new problem. Projects like Petals worked at it for years without ever reaching production. On HN, commenter "hazard" named the three fundamental issues — a single consumer card can't even hold one expert's weights, cross-node activation transfer slows each token to hundreds of milliseconds, and a P2P network has no reliable anti-cheat except duplicate computation (SHA-256 hashes files, not computation results). The README does not answer any of the three.
None. Pure open-source experiment. No hosted service, no pricing page, no revenue path.
Read: this question is premature — the project has not even reached the point of asking it. The more interesting question is that if P2P inference ever works, the money form would be incentives for node contributors (deposit-and-forfeit, like Bittensor), not software licenses. Too early to discuss.
| Dimension | Call |
|---|---|
| Founder-product fit | Genuine motivation (a natural extension of Colibrì), but one person with no datacenter limits what can be validated |
| Product insight | "Ship activations, not weights" is the right economy for MoE, and the Napster analogy explains it in one second; but it does not answer the three structural P2P problems |
| Execution quality | Full stack — encryption, signing, deterministic execution, DoS defenses, 80 commits in ten days. Serious engineering attitude |
| Timing | Too early, and path-dependent on an unproven assumption (strangers willing to run models for each other) |
This is a hypothesis about AI model distribution, not a product. Worth following, but graded as a hypothesis.
Three points worth keeping:
Why Unproven rather than Worth watching is straightforward: low attention (8 HN points), no business model, and none of the three structural problems addressed. But if in three months it demonstrates usable speed across an internet-scale swarm, the architecture notes deserve a careful read.
① Whether any independent third party (not the author) reproduces the speed numbers at real internet latency ② Whether any contributor incentive appears — P2P inference without incentives has never run in production ③ Whether stars and community engagement pick up — if they do not, even the developer community is signaling disbelief in the direction
Product logic: if you hold a compute-heavy task that is verify-able and read-heavy/write-light, outsourcing computation to a network is worth a serious look — cheaper than renting GPUs and cheap to verify. If your task is generative and subjective, cross P2P outsourcing off the list; anti-cheat costs will eat the gains.
Pricing structure: none. Not disclosed.
Unproven — logged with a probation period. A beautiful hypothesis, serious engineering, and no product destiny. If P2P inference as a direction is falsified, Lumabri is a high-quality technical notebook; if it is confirmed, this is the first complete blueprint. Come back in three months against the three checks above.
05
Use these searches when the official site is missing or the current link is only a lead.