Repository navigation
A relayed message is taken as its sender's only when the sender signed it. When two
computers cannot reach each other directly, a third forwards their requests and model work,
sealed so it cannot read them. The seal proved nothing about who made a message: another
computer connected to yours could send one naming someone else as its sender, and it was
handled as theirs. Now the sending computer signs each relayed message inside the seal, and
this version accepts one only when the signature checks, it was not delivered before, and it
is a kind of message computers relay. Direct connections were never affected. Relaying
between this version and an older one pauses until the older one updates itself, which takes
minutes; an older computer can still pass messages between two on this version.
Gemma 4's other sizes run. The 26B-A4B (a mixture of experts: 26B in total, 4B used per
token) loads, splits across computers and between a graphics card and the processor, and
answers as llama.cpp does: median cosine 0.99911 at 28 tokens and 0.99917 at 1,536 on the
processor, where two of llama.cpp's own processor code paths agree at 0.99895 and 0.99907. The
E2B and E4B (text) run whole on one computer — only their first layers see the prompt's words,
so they are not split across computers — and fit a small card: the E2B runs whole on a 4 GB
card and the E4B on a 6 GB one (card estimates 1,859 MB and 3,483 MB, where they were
3,696 MB and 5,688 MB).
A mixture-of-experts model that does not fit the graphics card reads prompts faster. Its
experts can now run on the card in the storage types Gemma 4 uses. A 1,527-token prompt to
Gemma 4 26B-A4B with every expert in the computer's memory read its three chunks in
4.0-10.3 / 3.2 / 3.1 s on the card, where it took about 10 s a chunk on the processor, and
23.6-24.8 s for the first chunk on the card.
A model split between the graphics card and the processor takes on more layers instead of
failing the request. A tester's 12 GB card refused a second run of Qwen3-30B-A3B's layers on
the card alone, with plenty of the computer's memory free, and the request failed; a whole copy
of the same model answered in 6.7 s.
A part fetched for one request is no longer deleted minutes later, and a model held whole
keeps every part. A tester's computer holding all 34 parts of Qwen3-30B removed part 25,
fetched it again and removed it 28 seconds after it arrived, three times in 13 hours. Trimming
now measures against the highest need of the last day, keeps every part of a model held whole
unless the disk is nearly full, and keeps a part that arrived in the last 30 minutes while
there is room.
Routes are priced from measurements that fit the prompt. A model load or a round trip timed
on a small forward and scaled up to a long prompt priced one route at 17.6 hours, so a fast
computer could lose to a slow one; the price now comes from warm forwards of at least 256 KB. A
prompt estimated at 1,024 tokens or more is counted exactly before it is planned (it was
estimated at 4 bytes a token). When this computer itself turned a request down, diagnostics no
longer blame the first peer in the plan.
A connection that has just dropped is no longer mistaken for a peer that stopped
answering. Twenty requests in flight on a closed connection read as "has not answered 20
requests in a row" — 14 of this computer's 16 such warnings on 2026-10-16. Only a timeout
counts now.
Less upload when computers reconnect. A computer connecting to a peer was sent all of the
peer's model descriptions, about 35 KB each: a test computer joining the network received 84 of
them (about 2.9 MB) from six peers in 20 seconds, describing 26 models. Between computers on
this version it is told which version of each the peer holds, about 200 bytes a model, and
fetches only what it lacks: about 12 KB per peer for those 26 models, and nothing for a peer
that already has them. The repeat to the whole network drops from every 2 hours to once a day
while every connected peer runs this version.
Five library patch updates (serde_json, toml, zeroize, interprocess, uuid), and the zip library
inside the model loader moves to 9.0.4 for three advisories filed on 9-10 October
(RUSTSEC-2026-0367, -0368, -0369). They concern unpacking an archive to disk, which SwarmLLM
never does: the library is there only for two model formats SwarmLLM does not read.
Full commit list: v0.3.240-alpha...v0.3.241-alpha