What works, what does not,
and what is next.

Every claim on this page has a measurement behind it. Where something was tried and failed, that is here too — a roadmap that only lists successes is advertising.

In italiano. Questa pagina dice cosa funziona davvero, cosa non funziona, e cosa manca. Ogni affermazione ha dietro una misura fatta su macchine vere. Anche le cose che abbiamo provato e che non hanno funzionato sono scritte qui: una roadmap che elenca solo i successi è pubblicità, non un documento.

The question we hold ourselves to

If neurionproject.org is switched off forever tomorrow, what keeps working?

Today the answer is everything, with one reservation: a brand-new install needs one introduction — any address that answers. eMule needed one too. After that, no list matters.

Built and proven

Models travel between people

If somebody has a model you want, it comes from them instead of from the internet. Checked against the publisher's own hash, so a bad copy costs a retry and nothing else — a hostile peer can waste your bandwidth and nothing more.

Measured: 986 MB between two machines; 398 MB relayed across the internet, hash identical.

A distributed index — no server in the middle

You can find a model on a machine you have never met, in a few hops through strangers. The node's identity is the fingerprint of its key, so it cannot choose where it sits; only nodes that answered you enter your routing table; an announcer's address is the one observed, never the one claimed.

Measured: a machine knowing one address found a model on a third it was never told existed.

Borrowing a processor, and checking the answer

A machine that cannot run a model can ask one that can. When two peers answer, both are asked to run reproducibly — temperature zero and the same seed — and the answers are compared.

Measured on two different processors: the answers came back identical, byte for byte. Before that fix, two honest peers disagreed at 0.45 word overlap and the check was worse than useless.

Reaching people who cannot be reached

Mobile networks and much shared fibre put whole neighbourhoods behind one address; no amount of asking the router changes it. Such a peer now attaches to one that is reachable and is contacted back through that connection — the arrangement eMule called a LowID.

Measured on a phone hotspot: unreachable directly, and 398 MB delivered anyway, correct hash.

Limits, consent, and a way to say no

a few transfers at a time, one per machine, under a rate ceiling you set;
traffic on your own network is exempt — the ceiling protects the line out of the house;
your router is asked, never opened silently;
one peer can be ignored without switching off sharing for everybody;
updates are signed with a key that has never been on the web server, so a compromised website cannot push code to your machine.

Three systems, and what each one actually gets

Windows, macOS and Linux installers are built from the tagged source by a machine nobody logs into. They are not the same in every respect, and pretending otherwise would waste your evening:

	Windows	macOS	Linux
Text, and everything the network moves	yes	yes	yes
Voice	yes	yes	yes
Video	yes	needs ffmpeg installed	yes
Images	yes	yes	Ubuntu 24.04 or newer

The two exceptions are somebody else's build, not our decision. There is no macOS ffmpeg from the project we take it from, so a Mac uses one you install — the app says so and finds it by itself. The image engine's Linux build is compiled on Ubuntu 24.04 and asks for a newer glibc than 22.04 has; where it cannot run, Neurion now says which glibc is missing instead of failing with a linker error after a 50 MB download.

Text was checked on the same old system, because that would have been the serious version of the problem: it needs glibc 2.34, runs on 2.35, and is unaffected. Everything the peer network carries works on all three.

Measured, and it changed our minds

Two beliefs this project held and had to drop when somebody actually measured them:

We believed	The measurement
Splitting a model across the internet is hopeless	False. Half here, half 32 ms away: 5.6 tok/s, against 5.4 for the far machine alone. The extra hop does not show up — the bottleneck is the slowest processor, not the network.
Routing between small specialists approaches a big model	False. Specialists are much better at home (+19 to +26 points) but no cheap router captures it on natural phrasing: measured +4 on unseen questions, inside the noise.

Deliberately not done

Splitting one computation across many machines. Possible, and slow: the activations cross the network every token, and a long prompt costs far more than the generation. Real work, not a promise.
Training models. Distributed training over home connections is a different problem and we are not pretending otherwise.
Being faster or cheaper than a commercial API. It is not. It is more independent, which is a different thing and worth saying plainly.
An installer signing certificate. Costs money we have not spent, so Windows shows an "unknown publisher" warning and macOS calls it an unidentified developer — on a Mac, right-click the app and choose Open the first time. Verify the SHA-256 against /download/latest.json, which is signed with a key that has never been on the web server; that part is real, and it is the part that actually protects you.

Hole punching. Today a relayed transfer passes through somebody else's connection. With hole punching the third party only makes the introduction and the bytes go direct. Needs a UDP transport.
A Kademlia-style DHT with no starting list at all. One introduction is still needed; Kad removed even that.
More entry points. Anyone can run one — no models, no compute, one small process and an open port. The fewer that matter, the healthier this is.

The honest bottleneck is not code. The network needs people. If you keep Neurion running with an open port, you are one of the ways in — and there is nothing to apply for.

What works, what does not,and what is next.