SN 53. Rank 14 of 32 by emission
engy
engy serves frontier open models on consumer GPUs, with cryptographic verification that the model you pay for is the model that ran.
In their own words · raw.githubusercontent.comSN53 engy sells answers from open AI models by the token, like a petrol station whose pump prints a receipt proving which grade you got, except the pumps are strangers' gaming cards.
The commodity, explained from zero
Inference is asking a trained AI model a question and getting an answer; it is sold by the token, roughly a word. engy sells that for seven open models, priced per million tokens on the front page, from 0.045 dollars in to 9.75 dollars out. Billing is "Per-token, pay as you go. No subscriptions and no minimums."
The twist is the receipt. Anyone selling inference can quietly run a cut-down copy of a model and bill for the full one, the way a station could pump regular and charge for premium. engy's answer is a proof with every response: the model's weights are pinned by a published fingerprint (a Merkle root), and each reply carries a compact summary of the model's internal activity (TOPLOC) that a checker can compare against the real model. The README's phrase is that it "pins the math, not the machine."
The buyers named are AI agents and their builders; the docs give setup guides for Claude Code, Cursor, and Codex, and any OpenAI-compatible client can point at api.engy.ai/v1. No customer is named and no usage figure appears.
Why it is on Bittensor at all
The supply comes from consumer hardware: the reference miner setup is four RTX 4090 gaming cards, and the status page today lists 227 providers for the smallest model. That is what the petrol-station comparison breaks on: a station owns its pumps, while engy's pumps belong to strangers who dial in over the internet and are paid in the subnet's token. The proof exists because of that: the README says "No TEEs, no trusted hardware", so a stranger's card is trusted on the maths, not on a locked-down chip. Cost against a centralized provider is not compared on the site.
How the work gets done
Miners run an open model on their own GPUs, take requests from engy's gateway, and return each answer with its proof. Validators, the judges who decide pay, sample requests and check the proofs; in the team's own test a cut-down model scores about 0.55 where an honest one scores about 0.2. Each week a miner passes four gates, 99 percent of requests accepted, two speed limits, and at most 1 percent cheat verdicts, or earns nothing that week; those that pass are paid in proportion to paid tokens served, and only paid traffic counts, so miners cannot buy their own requests. Emission, the subnet's share of new tokens, follows those weights.
How you would know it works
The status page at engy.ai/status shows each model, its provider count, and ninety-day uptime (99.71 percent for the gateway). A reader without an API key cannot try the model, and the cheat-detection experiment has no date.
What is missing
No usage or revenue figure exists on any page. There is no staker page, and the chain identity has no Discord field though the site links one. The cheat-detection numbers live in a technical one-pager, and the site never explains "verified inference" in plain words. Registration cost is not stated. There is no customer SLA or escalation path beyond email and Discord.
Go deeper
Sources for this explainer
Metaphor: a petrol station whose pump proves the grade. Every claim is drawn from the evidence set or the subnet's own materials; "(inferred)" marks a conclusion rather than a quote. Corrections.
How we scored it
Four audiences, four questions each, scored on what a first-time reader can find in five minutes. Method in the rubric.
- subnet_name
- engy
- github_repo
- github.com
- subnet_contact
- ning@engy.ai
- subnet_url
- engy.ai
- discord
- not set
- description
- Verified inference for frontier open models.
- additional
- not set
Stakers and validators●●●●● 3.5
Should I allocate here? · rank 9 of 32 for this audience
Q1What it is●●●●● 5
The fuel for a world run by agents.
with a per 1M token price table for seven models on the front page, README opens Verified inference for frontier open models which is the chain description word for word, unit is tokens
Q2Who it is for●●●●● 2
qwen3.6-35b-a3b | 227 providers | operational | 99.90% uptime
is supply and uptime, no usage, revenue or customer figure on any own page, a Buyback and Burn section sits behind the provider login
Q3How it resists gaming or fails●●●●● 4
one pager linked from the README names four scoring gates (acceptance 99 percent, two latency SLAs, cheat verdict rate 1 percent), anti wash trading (only paid 2xx traffic scores), an INT4 versus FP8 cheat experiment (0.55 against 0.18) and the rule that scoring never bans, technical register
Q4Identity and documentation●●●●● 3
identity 5 of 7, discord and additional not set though the site links a Discord invite, github and url resolve, contact ning@engy.ai in the README matches the chain, VALIDATOR.md describes a relay only light validator, no staker page
Miners●●●●● 3.8
Can I compete, and what wins? · rank 11 of 32 for this audience
Q1What it is●●●●● 4
serves buyer requests routed from a gateway on a local sglang serve and returns, with each completion, a compact TOPLOC proof
requests arrive over a WebSocket from the gateway, no request or response schema shown
Q2Who it is for●●●●● 3
4 × RTX 4090 (24 GB), Qwen3.6-35B-A3B-FP8
as the reference config and MAX_INFLIGHT minimum 8, a miner key comes from provider.engy.ai behind a login, chain registration cost not stated (burn 0.0005 TAO today), competitiveness stated as gates in the separate one pager
Q3How it resists gaming or fails●●●●● 4
weight formula floor of tokens times score_rate per model, four gates, a fresh key starts at zero, a miner that trips a gate earns nothing that epoch and recovers the next, cheat experiment numbers, split between the one pager and MINER.md and written for engineers
Q4Identity and documentation●●●●● 4
install guide of about five minutes plus model load with env settings, latest release v0.4.1 dated 2026-07-21, repo pushed 2026-09-03, guide undated
Buyers and enterprises●●●●● 4.5
Can I use this today? · rank 2 of 32 for this audience
Q1What it is●●●●● 5
Point any OpenAI-compatible client at https://api.engy.ai/v1 with your API key.
with integration guides for Claude Code, Cursor and Codex, seven models priced on the front page, agent API reference with a curl example
Q2Who it is for●●●●● 4
Per-token, pay as you go. No subscriptions and no minimums
per model input, output and cached prices, the customer named is agents and their builders, no enterprise tier or named customer segment
Q3How it resists gaming or fails●●●●● 4
All systems operational
with 90 day uptime per model and for the API gateway, email contact and a Discord invite in the docs footer, no customer SLA and no escalation path stated
Q4Identity and documentation●●●●● 5
agent API reference with base URL engy.ai/api/v1, bearer auth, 120 requests per minute per token and an error envelope, docs page for the inference endpoint, README contact ning@engy.ai matches the chain
Newcomers●●●●● 4.0
What is this and why does it matter? · rank 5 of 32 for this audience
Q1What it is●●●●● 4
engy serves frontier open models on consumer GPUs, with cryptographic verification that the model you pay for is the model that ran.
one sentence at the top of the README, the front page headline is a slogan, verified inference is left unexplained on the site
Q2Who it is for●●●●● 4
No TEEs, no trusted hardware.
plus weights pinned by a published Merkle root and a TOPLOC fingerprint per response, a concrete difference (proof the paid model ran) without a side by side comparison against a centralized provider
Q3How it resists gaming or fails●●●●● 4
90 day uptime history per model on the status page, an INT4 versus FP8 detection experiment with numbers in the one pager but undated, no public demo without an API key
Q4Identity and documentation●●●●● 4
subnet_name engy matches the site, description is a full sentence, url resolves, docs, pricing and status pages but no about page
How this score came to be. Verified means the scorer fetched the page and the words are on it; inferred means concluded from code, absence, or a third party. The link badge is a separate automated check made before publication.
Something wrong? Corrections of fact are applied as they arrive during the window; score disputes are batched at its close. How to file one · GitHub issue · email.