// Shard // Aleph Hackathon 2026 // Built on Pear //
sha/rd
A language model, sharded layer‑by‑layer across peers. On Pear.
§ 01 — Someone else holds the switch
25
countries where hosted assistants like ChatGPT are unavailable
ChatGPT availability — World Population Review, Aug 2026
313
internet shutdowns across 52 countries in 2025 — a record
Access Now, #KeepItOn

A model reachable at an address can be blocked at that address. Shard has no address.

§ 02 — How it works
laptop1
ggml-rpc-server :50053 · Metal
layers 0–14 · offers 1 GB
laptop2
ggml-rpc-server :50052 · Metal
layers 15–27 · offers 1 GB
Llama 3.2 3B Q4_K_M · layer 00layer 27
Discovery
UDP sweep on the LAN
ficha: label, RAM, GB offered, rpc port
Assignment
No coordinator
every node sorts peers by id and reaches the same split
Transport
llama.cpp RPC over TCP
one --rpc list drives every shard
Delivery
Pear · Hyperdrive
install by key, OTA deltas to running apps
Today LAN only · 2 peers · one owner per layer · answers with the internet cut Next WAN · redundancy
§ 03 — Pear-native
pear install --to ~/shard
pear://k6c99su98pmobmw1c4xqtoacsage1is5ayhq9mqdsc8gobdzg8bo

No registry, no release page. It arrives from the peers, and it updates itself over the air. You will see that happen in the demo.

§ 04 — What is ours, what is not

We built

  • Shard planning with no coordinator
  • Peer discovery and liveness on the LAN
  • Each node runs its own layer server
  • The status panel, over SSE

The stack gives

  • Pear — distribution, OTA, Hyperdrive
  • Bare — the standalone binary
  • llama.cpp — RPC backend, layer split

It does not

  • Prove a peer really ran your layers
  • Survive a peer leaving — no redundancy yet
  • Cross subnets, or use phones as shards
Enough slides.
Watch it think.
MODEL INCOMPLETE — LAYERS 15-27 MISSING 15 of 28 layers. The model cannot think until another machine joins.
01 / 06
← → · space · F