Swarm

SN124 Swarm is a pilot's exam for drone-flying software, where the course is redrawn every two weeks so nothing can be memorized, and the best pilot in each event holds the title until someone flies better.

The commodity, explained from zero

The README opens plainly: "Drones are moving into everyday life: delivering packages, inspecting bridges and power lines, searching for survivors after a disaster." Swarm's commodity is that software: a trained flight policy, a file that reads a single depth camera and the drone's position and speed and decides, fifty times a second, where to go, with "no map, no GPS, no list of obstacles."

The policy is trained and tested in a simulator of a Crazyflie, a small research quadcopter. There are five events, each with its own champion and share of the pay: Office Interceptor (30 percent), Swarm Autopilot (20), Swarm Search and Rescue (20), Autopilot (15), and Search and Rescue (15). Every fourteen days the simulator generates 1,100 fresh worlds.

Who pays: nobody yet, in the materials. The site calls itself "the lab for autonomous drone AI," and, rendered with JavaScript, lists sectors such as defence, search and rescue, and agriculture and promises "Model adapters for your airframes, sensors, and constraints," but names no customer, price, or way to buy. Here the comparison breaks: a pilot who passes keeps the licence, while a Swarm champion is dethroned the moment someone flies better.

Why it is on Bittensor at all

The README's argument is comparison in the open: "today that AI is built behind closed doors, where no one can compare it or prove whose is best." An open arena with unseen worlds and a public leaderboard lets anyone check the claim. What the network changes is who can enter and who can see the scores; cost and coverage are not claimed.

How the work gets done

Miners train a policy by whatever method they like, on any hardware; the miner process itself needs only 3 cores and 8 GB of RAM and "commits your model's digest on-chain, uploads the archive to the Swarm backend, and exits." Validators fly each policy through the fresh worlds and score it: 0.45 for landing successfully, 0.45 for time, 0.10 for safety, with a failed run scoring 0.01. A new champion must beat the old one by at least 0.015, a floor that shrinks toward 0.005 as scores approach perfect, so a copy with a tiny tweak earns only a tiny share. Each event's emission, its slice of newly minted TAO, is split across "its last five crowned kings," each older seat holding 70 percent of the seat above it.

How you would know it works

The benchmark page at https://swarm124.com/benchmark is the artifact, but it renders only with JavaScript; the index's fetch on 2026-09-09 saw champion scores per event and one event crowned on 2026-08-21.

What is missing

The chain identity is 4 of 7: no description, no Discord, no additional field, and the URL is set without a scheme. The contact admin@swarm124.com appears nowhere on the site or in the repository. There is no API reference, no pricing, and the only documentation is for miners. There are no GitHub releases.

Go deeper

• README with plain explanation: https://raw.githubusercontent.com/swarm-subnet/swarm/main/README.md
• Miner guide with scoring formula: https://raw.githubusercontent.com/swarm-subnet/swarm/main/miner/docs/miner.md
• King of the Hill document: https://raw.githubusercontent.com/swarm-subnet/swarm/main/docs/king_of_the_hill.md

Metaphor: a pilot's exam where the course is redrawn every two weeks. Every claim is drawn from the evidence set or the subnet's own materials; "(inferred)" marks a conclusion rather than a quote. Corrections.