With the amount of new subnets being added it can be hard to get up to date information across all subnets, so data may be slightly out of date from time to time

Subnet 97

Ingot

Subnet stats · Subnet

● Loading…
Alpha price
–
 
Market cap
–
 
Price (τ) & daily volume
Loading price history…
Daily emission split
Pool depth (TAO vs alpha value)
24h trading flow
What do these numbers mean?
Alpha price
TAO per alpha token, calculated from the subnet's on-chain pool (TAO reserve ÷ alpha reserve).
Market cap
Price × total alpha issued (alpha in pool + alpha outstanding). FDV = price × 21M max supply.
TAO liquidity
TAO currently held in the subnet's pool.
Alpha outstanding
Alpha held outside the pool (staked to validators/miners/owner).
Alpha emission / day
Alpha emitted to participants each day (on-chain), valued in TAO at the current price.
TAO injected / day
TAO the network adds to this subnet's pool each day (on-chain).
Net TAO flow / day
Smoothed (EMA) net TAO staked minus unstaked per day.
Neurons
Registered UIDs / maximum UIDs on this subnet.
Neuron reg. cost
Current on-chain burn (TAO) to register one miner/validator UID on this subnet. This is not the cost of creating a subnet.
Sources: Bittensor chain (Finney, direct read) · tao.com · TaoMarketCap · TAO/USD: CoinGecko

ABOUT

What exactly does it do?

Ingot is Bittensor Subnet 97, a model-compression competition summed up in its tagline “Open models, cast small.” A fixed open-source teacher model, currently Qwen3.8-27B in BF16, is the reference. Miners submit quantized FP8 copies of it with the same architecture and no larger than the subnet’s genesis model, and the validator crowns the copy that stays closest to the teacher. Every crowned model is published openly on Hugging Face.

The subnet was previously known as Albedo (and before that distil), a distillation subnet run by the same team. On 16 October 2026 it was renamed Ingot with a new mechanism: instead of distilling a large model into a smaller one, miners now compete to produce the most faithful compressed (quantized) version of the teacher. The old albedo.tech site now points visitors to theingot.ai.

How a submission is judged

  • Fixed reference: the teacher answers prompts drawn from a large multilingual evaluation dataset; the reigning king and the challenger then read exactly the same prompt and answer, and only the answer tokens are scored.
  • Two measurements per token: KL divergence between the teacher’s and the model’s next-token distributions, and how far the model’s hidden states drift from the teacher’s across a set of layers. Lower is better on both.
  • King of the hill: a challenger must beat the current king by a margin on both measures to be crowned. A win makes it the new king immediately; a loss is final for that hotkey.
  • Original work only: before evaluation the validator fingerprints each submission’s quantization error and rejects near-exact copies of genesis or of other miners’ earlier models.

Who gets paid

The newest five kings each receive 20% of miner emission for five days after their coronation. Whatever is not assigned to a reigning king is burned to UID 0, so with a single king 20% goes to it and 80% is burned, while five fresh kings share the full amount.

Ingot is Bittensor Subnet 97, a model-compression competition summed up in its tagline “Open models, cast small.” A fixed open-source teacher model, currently Qwen3.8-27B in BF16, is the reference. Miners submit quantized FP8 copies of it with the same architecture and no larger than the subnet’s genesis model, and the validator crowns the copy that stays closest to the teacher. Every crowned model is published openly on Hugging Face.

The subnet was previously known as Albedo (and before that distil), a distillation subnet run by the same team. On 16 October 2026 it was renamed Ingot with a new mechanism: instead of distilling a large model into a smaller one, miners now compete to produce the most faithful compressed (quantized) version of the teacher. The old albedo.tech site now points visitors to theingot.ai.

How a submission is judged

  • Fixed reference: the teacher answers prompts drawn from a large multilingual evaluation dataset; the reigning king and the challenger then read exactly the same prompt and answer, and only the answer tokens are scored.
  • Two measurements per token: KL divergence between the teacher’s and the model’s next-token distributions, and how far the model’s hidden states drift from the teacher’s across a set of layers. Lower is better on both.
  • King of the hill: a challenger must beat the current king by a margin on both measures to be crowned. A win makes it the new king immediately; a loss is final for that hotkey.
  • Original work only: before evaluation the validator fingerprints each submission’s quantization error and rejects near-exact copies of genesis or of other miners’ earlier models.

Who gets paid

The newest five kings each receive 20% of miner emission for five days after their coronation. Whatever is not assigned to a reigning king is burned to UID 0, so with a single king 20% goes to it and 80% is burned, while five fresh kings share the full amount.

PURPOSE

What exactly is the 'product/build'?

What exists today

The live product is the Ingot competition and the open models it produces. Every king is mirrored to Hugging Face under huggingface.co/dendriteholdings as Ingot-<roman numeral>-27B-FP8, published together with the genesis tokenizer and configuration files so anyone can download and run it. The starting point, Ingot-genesis-27B-FP8, was crowned on 16 October 2026, and by the morning of 17 October miners had already produced three kings (Ingot-I, Ingot-II and Ingot-III) from 26 evaluated submissions.

Dashboard

theingot.ai is the subnet’s live dashboard. It shows the reigning king and its Hugging Face repository, a chart of every coronation, the current weight holders and when their reigns end, registration activity, the evaluation queue, a full duel history, and failed submissions with their reasons.

Mining workflow

  • Miners use the open-source ingot command-line tool to register, run local validity checks (ingot check-model) and submit.
  • Models are uploaded privately: the validator issues one-time storage credentials scoped to the miner’s own folder, and a model only becomes public if it wins.
  • Submissions must keep every genesis tensor’s name and shape, use FP8 (E4M3) where genesis does, and stay within the genesis size of about 30.9 GB. One submission is allowed per hotkey.

Open data and code

The evaluation prompts come from the public Hugging Face dataset dendriteholdings/ingot_corpus, a multilingual corpus of more than 11 million prompts across dozens of languages, sampled per domain and seeded by the block hash of each miner’s commitment so every validator draws the same set. The validator, evaluator, miner tooling and website are published at github.com/DendriteLabs/ingot.

What exists today

The live product is the Ingot competition and the open models it produces. Every king is mirrored to Hugging Face under huggingface.co/dendriteholdings as Ingot-<roman numeral>-27B-FP8, published together with the genesis tokenizer and configuration files so anyone can download and run it. The starting point, Ingot-genesis-27B-FP8, was crowned on 16 October 2026, and by the morning of 17 October miners had already produced three kings (Ingot-I, Ingot-II and Ingot-III) from 26 evaluated submissions.

Dashboard

theingot.ai is the subnet’s live dashboard. It shows the reigning king and its Hugging Face repository, a chart of every coronation, the current weight holders and when their reigns end, registration activity, the evaluation queue, a full duel history, and failed submissions with their reasons.

Mining workflow

  • Miners use the open-source ingot command-line tool to register, run local validity checks (ingot check-model) and submit.
  • Models are uploaded privately: the validator issues one-time storage credentials scoped to the miner’s own folder, and a model only becomes public if it wins.
  • Submissions must keep every genesis tensor’s name and shape, use FP8 (E4M3) where genesis does, and stay within the genesis size of about 30.9 GB. One submission is allowed per hotkey.

Open data and code

The evaluation prompts come from the public Hugging Face dataset dendriteholdings/ingot_corpus, a multilingual corpus of more than 11 million prompts across dozens of languages, sampled per domain and seeded by the block hash of each miner’s commitment so every validator draws the same set. The validator, evaluator, miner tooling and website are published at github.com/DendriteLabs/ingot.

WHO

Team Info

Team

Ingot is built by Dendrite (Dendrite Holdings, dendrite.holdings), a Bittensor company dating from 2022 and led by its founder and CEO Piotr Barbachowski. Dendrite mines across Bittensor, owns Subnet 114 SOMA, and co-develops Subnet 3 Teutonic, Subnet 66 Conjectures and Subnet 97, which it previously co-developed as Albedo. Dendrite describes itself as co-developing subnets with Bittensor co-founder Jacob “Const” Steeves.

Ingot’s lead developer is Weather Boss (X @ingot_wthrboss, formerly the Albedo updates account), whose GitHub account weather-boss-dendrite authors the Ingot codebase under the DendriteLabs organisation.

Subnet ownership

The rename to Ingot kept Subnet 97’s existing registration (block 7,735,450) and owner coldkey, so this is a rebrand of the same subnet rather than a new registration.

Community

Updates are posted by @ingot_wthrboss and Dendrite (@DendriteHQ). No dedicated Ingot Discord invite is published, so community discussion defaults to the main Bittensor Discord.

Team

Ingot is built by Dendrite (Dendrite Holdings, dendrite.holdings), a Bittensor company dating from 2022 and led by its founder and CEO Piotr Barbachowski. Dendrite mines across Bittensor, owns Subnet 114 SOMA, and co-develops Subnet 3 Teutonic, Subnet 66 Conjectures and Subnet 97, which it previously co-developed as Albedo. Dendrite describes itself as co-developing subnets with Bittensor co-founder Jacob “Const” Steeves.

Ingot’s lead developer is Weather Boss (X @ingot_wthrboss, formerly the Albedo updates account), whose GitHub account weather-boss-dendrite authors the Ingot codebase under the DendriteLabs organisation.

Subnet ownership

The rename to Ingot kept Subnet 97’s existing registration (block 7,735,450) and owner coldkey, so this is a rebrand of the same subnet rather than a new registration.

Community

Updates are posted by @ingot_wthrboss and Dendrite (@DendriteHQ). No dedicated Ingot Discord invite is published, so community discussion defaults to the main Bittensor Discord.

FUTURE

Roadmap

Ingot has not published a dated roadmap. Its repository and documentation point to the following direction:

  • Stricter copy detection: the duplicate-detection thresholds are deliberately loose at launch, and the documentation lists stricter calibrated values (0.7 and 0.8) as the intended setting.
  • More model families: the upload rules are organised by model family and quantization pair, with only qwen38_27b:fp8 listed today, so further teachers or formats can be added through the same configuration.
  • A growing line of open kings: each coronation adds another openly published FP8 model to the dendriteholdings collection on Hugging Face.

Changes to the mechanism are published first in the repository’s chain.toml and configuration files, which the project describes as the source of truth for what miners and the validator must agree on.

Ingot has not published a dated roadmap. Its repository and documentation point to the following direction:

  • Stricter copy detection: the duplicate-detection thresholds are deliberately loose at launch, and the documentation lists stricter calibrated values (0.7 and 0.8) as the intended setting.
  • More model families: the upload rules are organised by model family and quantization pair, with only qwen38_27b:fp8 listed today, so further teachers or formats can be added through the same configuration.
  • A growing line of open kings: each coronation adds another openly published FP8 model to the dendriteholdings collection on Hugging Face.

Changes to the mechanism are published first in the repository’s chain.toml and configuration files, which the project describes as the source of truth for what miners and the validator must agree on.

NEWS

Announcements

Latest on X

Loading…