SOMA is a Bittensor subnet focused on context compression for AI agents.
SN 114. Rank 10 of 32 by emission
SOMA
Emission2.4%of the block reward · rank 10 of 32
Participants4 miners · 9 validators
Registration0.25 TAO
Verified2026-09-09links checked and scores merged
The programinference · github.com@83b828d · read 2026-09-19
The work
Compression code. A miner submits an algorithm, model or hybrid that shrinks the context a coding agent sends to a language model: conversation history, tool output, repository files. The miner runs no server; the platform and validators run the agent with and without the compressor. What is paid for is the compressor that saves the most weighted tokens while the agent still solves its tasks.
own docs · raw.githubusercontent.com
SOMA is a Bittensor subnet focused on context compression for AI agents.
How it is scored
Per task, x is how many baseline runs solved it and y how many miner runs did. The token score is log2(baseline tokens / miner tokens) clamped to [-2, 2], output tokens counted at 3, cached input at 0.1. At y/x of 0.5 or under the score runs from -2 to -1; from 0.5 to 0.8 it blends toward the token score; above, token score plus a bonus to 0.1.
code · raw.githubusercontent.com · mcp_platform/app/api/routes/scoring.py
SCORING_BONUS_MAX = 0.1
Sources disagree docs/miner/scoring.md at the pinned commit, mirrored at thesoma.ai/docs/miner/scoring, gives a bonus cap of 3.0, a penalty floor of -4.0 and a threshold of floor(0.8x); mcp_platform/app/api/routes/scoring.py at the same commit caps the bonus at 0.1, starts the penalty at -2 and uses ceil(0.8x).
How the pay splits
Winner takes each contest. Tasks are classed short, medium or long; the three-way contest carries 0.25 of the miner weight, the three pairs 0.15 each, the three singles 0.10 each. The top scorer in each contest takes its weight, ties split evenly, and a miner with no score in a member class is not in that contest. Whatever the platform's burn ratio withholds goes to burn.
code · raw.githubusercontent.com · mcp_platform/app/services/incentive_calculator.py
LAYER_WEIGHTS_BY_SUBSET_SIZE: dict[int, float] = {3: 0.25, 2: 0.45, 1: 0.30}
Who judges
Nine validators were active on the 2026-09-09 chain snapshot, of 64 permitted. Commit-reveal is off; weights may change at most once per 100 blocks, and the code sets them every 110. A validator runs agent tasks and grades patches in a container, but the weights it submits arrive signed from the platform's get_best_miners endpoint, with a cached copy as fallback. The burn ratio is a platform database setting.
chain · hyperparams-2026-09-19.json
commit_reveal_weights_enabled False, weights_rate_limit 100, max_validators 64; chain-2026-09-09.json active_validators 9
How often it turns
Emission settles every 360 blocks, about 72 minutes, and weights are set every 110 blocks. The puzzle turns per competition, not per epoch: an upload window, an idle gap, then a screening stage and full evaluation on hidden tasks whose images go public only for the evaluation window plus a grace period. A new registration is immune from deregistration for 5,000 blocks, about 16 hours.
chain · hyperparams-2026-09-19.json
tempo 360, immunity_period 5000
Who buys
Developers routing model calls through SOMA. The front page prices one model, DeepSeek V4 Pro, at $1.32 per million tokens direct and $1.19 through SOMA, about 10 percent off, with two more models marked coming soon. The tools listed are SOMARIZER, a Compressor for OpenClaw and a SOMA App. No customer is named and no dated usage figure appears on the site.
own docs · thesoma.ai
$1.32 / 1M $1.19 / 1M ~10% off
What has been gamed
Prompt edits that change the agent instead of compressing it: the miner rules allow only compression markers and loop guards, and forbid rewriting instructions or forcing shortcuts to cut tokens. Memorizing the test: evaluation tasks live in a private image repository made public only for the evaluation window. Blind deletion: a miner at or below half of baseline quality scores the penalty alone. Top submissions get an integrity review.
code · raw.githubusercontent.com · miner/README_prompting.md
Prompt edits that force shortcuts to reduce token usage.
5 of 7 fields rest on the chain or the code. A manifest is the site's reading of the subnet's program, not a review; each field links its source. All programs · What a field means · Correct a field
What it says
The subnet in its own words: the line the index took from its front door, and what the team has said on the record.
What it shows
Explained from zero by the index, from the subnet's own materials, for someone who has never heard of Bittensor.
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
Sources for this explainer
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.
How legible it is
Four audiences, four questions each, rated on what a first-time reader can find in five minutes. The scale measures findability, not quality.
On-chain identity 6/7- subnet_name
- SOMA
- github_repo
- github.com
- subnet_contact
- x.com
- subnet_url
- thesoma.ai
- discord
- discord.gg
- description
- Context compression layer delivered through MCP infrastructure
- additional
- not set
Stakers and validators●●●●● 3.8
Should I allocate here? · rank 6 of 32 for this audience
Q1What it is●●●●● 4
Cut Token Costs Keep The Quality
with a per 1M token price for one model with and without SOMA, README says a subnet focused on context compression for AI agents, matches the chain description, unit is tokens saved rather than a stated commodity so took the lower score
Q2Who it is for●●●●● 3
Save Tokens
chart of daily token savings up to 220k over five days on the front page, no date, no named customers, no revenue statement
Q3How it resists gaming or fails●●●●● 4
named incentive page dated 15 Jul 2026 two clicks from the docs root with benchmark weights 50/25/25, a layered winner takes element share formula and a burn ratio, no anti-gaming section and no failure modes, a separate prompt editing rules page holds the miner rules
Q4Identity and documentation●●●●● 4
identity 6 of 7 (additional not set), github, url and discord fields are links, subnet_contact is the X account which the contact page also lists, no staking page but a SOMA Lock Program page states terms for locking alpha
Miners●●●●● 3.0
Can I compete, and what wins? · rank 25 of 32 for this audience
Q1What it is●●●●● 3
build a solution that reduces the amount of context processed by the agent while preserving its ability to successfully complete the underlying task
miner setup page says args and return types vary by competition, the compression page (Feb 2026) gives injected text plus a ratio in and compressed text out, no schema
Q2Who it is for●●●●● 2
4 CPU cores, 16 GB RAM, 500 GB SSD storage
appears under the Validators heading, no miner hardware statement, registration command given with no cost (chain burn 0.25 TAO today), no competitiveness statement
Q3How it resists gaming or fails●●●●● 4
scoring page with weighted token formula, compression ratio clamp, penalty floor and constants table, incentive page gives winner per element shares, no worked example, no immunity or deregistration policy
Q4Identity and documentation●●●●● 3
miner setup guide last updated 27 May 2026, release v1.0.0 2026-05-20, repo pushed 2026-09-09, incentive page changed 15 Jul 2026 after the guide, guide defers to per competition requirements
Buyers and enterprises●●●●● 3.5
Can I use this today? · rank 10 of 32 for this audience
Q1What it is●●●●● 4
Official SOMA application for AI agents
plus SOMARIZER and a Compressor for OpenClaw listed as tools on the front page, app.thesoma.ai is a JS dashboard, no API endpoint or example found
Q2Who it is for●●●●● 4
AI teams are paying for repeated context, not better intelligence
with Deepseek V4 Pro at $1.32 / 1M against $1.19 / 1M with SOMA, other models marked coming soon, no pricing page beyond the front page
Q3How it resists gaming or fails●●●●● 3
contact form, thesoma@dendrite.holdings and a Discord invite on the contact page, no status page (status.thesoma.ai does not resolve), no SLA, no response time
Q4Identity and documentation●●●●● 3
no API reference in the docs which cover miners, validators and benchmarks only, the contact page lists https://x.com/SomaSubnet which matches the chain subnet_contact
Newcomers●●●●● 3.3
What is this and why does it matter? · rank 20 of 32 for this audience
Q1What it is●●●●● 4
a compression layer that cuts token costs without cutting output quality
on the front page, README opens with a one sentence definition, consistent with the chain description
Q2Who it is for●●●●● 2
AI teams are paying for repeated context, not better intelligence
frames the difference as cost versus redundant context, nothing on what decentralization changes or a comparison with a centralized compressor
Q3How it resists gaming or fails●●●●● 3
token savings chart on the front page without a date, a live SOMARIZER tool, benchmark docs describe how to run SOMA Bench with no published results table
Q4Identity and documentation●●●●● 4
subnet_name SOMA matches the site, description is a noun phrase without a finite verb, url resolves, docs and contact pages serve as about and learn pages
Provenance
How this rating came to be. Verified means the scorer fetched the page and the words are on it; secondhand means concluded from code, absence, or a third party. The link badge is a separate automated check made before publication.
scored by agent b 2026-09-09, rubric v1.0merged 2026-09-09links verified 2026-09-09: 32 live, 0 unreachable, 1 manual
Something wrong? Corrections of fact are applied as they arrive during the window; rating disputes are batched at its close. How to file one · GitHub issue · email.