SOMA

SN114 SOMA is a packing contest for AI agents, where miners compete to fit the same conversation into a smaller suitcase so the model reads fewer words and the bill comes down without the answer getting worse.

The commodity, explained from zero

An AI agent is a program that asks a language model for the next step, over and over, sending it the whole conversation so far each time: the task, the files it opened, the commands it ran. Models charge per token, roughly per word, for everything they read, so an agent that works for an hour pays again and again to re-read its own history. The README says it plainly: "Modern AI agents repeatedly send large amounts of context to language models."

SOMA sells the packing: a compressor sits in front of the model and shortens the context. The unit of value is tokens not sent. The front page shows the price of one model, DeepSeek V4 Pro, at 1.32 dollars per million tokens without SOMA and 1.19 with it; four other models are marked "coming soon". A chart shows daily savings of up to about 220,000 tokens over five days, with no date.

The stated buyers are "AI teams" that are "paying for repeated context, not better intelligence". Three tools are listed: SOMARIZER for documents, a compressor for the OpenClaw coding agent, and a SOMA app. No customer is named.

Why it is on Bittensor at all

Decentralization here means the compressor is chosen by open contest rather than written in-house. Anyone can submit a compression method, and the network tests it on real coding tasks against the uncompressed version; the README describes the goal as "the maximum useful compression an AI agent can sustain while continuing to complete real work". On cost, ownership, or access against a centralized compressor, the materials say nothing yet. The comparison breaks here too: a packing contest judges the suitcase, but SOMA judges the trip, scoring whether the agent still finished the job.

How the work gets done

Miners submit compression code; they do not run servers, and the task's inputs and outputs change from one competition to the next. Validators, the judges who decide pay, run standard software-engineering benchmarks (SWE-bench Verified and two exploration and editing variants, weighted 50, 25 and 25 percent) with and without each miner's compressor, then score the trade between tokens saved and tasks still solved; output tokens count three times input, and ratio credit is capped at fourfold (read from the scoring formula). The winner of each benchmark combination takes that slice of emission, the subnet's share of new tokens, and a set portion is burned. The incentive page is dated 15 July 2026.

How you would know it works

The front page's savings chart and price comparison are the only artifact, and the chart carries no date. The docs describe SOMA Bench, a benchmark a reader can run, but publish no results table. SOMARIZER can be tried.

What is missing

There is no API reference; the docs cover miners, validators, and benchmarks only. The miner setup guide is from May 2026 and the incentive rules changed in July. Hardware is stated for validators, not miners, and registration cost is not given. There is no status page or SLA; support is a contact form, an email, and Discord. Nothing states how a miner is dropped.

Go deeper

• Front page with price comparison
• README explaining the problem
• Incentive mechanism, July 2026

Metaphor: a packing contest for AI agents. Every claim is drawn from the evidence set or the subnet's own materials; "(inferred)" marks a conclusion rather than a quote. Corrections.