SN 97. Rank 6 of 32 by emission
Albedo
King-of-the-hill trajectory-distillation subnet on Bittensor (netuid 97, finney).
In their own words · albedo.techSN97 Albedo is a playground game of king of the hill for coding models, where one fine-tuned model stands on top and earns most of the prize until a challenger knocks it off by a clear margin, twice.
The commodity, explained from zero
The output is a model checkpoint: a saved copy of an AI model's weights, the file needed to run it. Every entrant climbs from the same flat ground, the open model Qwen3.6-35B-A3B, and fine-tunes it, meaning they train it a little further on their own data for one job. The job here is acting as a coding agent: reading a real software repository, running commands, and making edits over several turns. The materials call the recorded runs "trajectories" and the method "trajectory distillation", a term they do not define (inferred from the scoring doc: learning from example runs of a stronger model).
A better coding-agent model could have a price, but Albedo's materials name no buyer, no price, and no unit of sale. The only public use is a "Chat with the King" link to the current champion. For someone to pay, the champion would have to beat the models they already rent; the site's benchmark panel should show that but rendered only "loading" when fetched.
Why it is on Bittensor at all
The honest answer is: an open contest with a public ledger. Anyone can register, upload a fine-tuned model to a file host, and commit a fingerprint on chain; the king is chosen by a fixed procedure rather than a lab's release schedule. The materials do not compare this with a centralized provider on cost, ownership, or censorship. Here the comparison breaks: on the playground, holding the hill is the whole point, but here the model on top is supposed to be the product, and nobody has said who wants it.
How the work gets done
Miners fine-tune the base model, keep its configuration files identical to the original, and publish it; a duplicate check rejects copies and lightly noised copies of earlier entries. Validators, the judges who set pay, run the challenger and the reigning king on the same hundred coding tasks from real repositories, then ask a judge model (GLM 5.2) a checklist of yes-or-no questions about each run; a run that loops on one command scores zero. A challenger takes the hill only if it beats the king by 2.5 points in two separate evaluations. Emission, the subnet's share of new tokens, is split across the current king and the previous four in a rolling chain, so a king pushed off the hill fades out rather than dropping to zero.
How you would know it works
No artifact a reader can check exists today, in those words. The dashboard at albedo.tech lists benchmarks, reign, queue, history, and fails, but each section shows "loading…" to a fetcher, and the chat link opens a chat app that explains nothing.
What is missing
The chain description is two words, "Alchemical intelligence". The repository has no README; the guides sit in a docs folder nothing links to. The chain contact appears on no page. There is no price, customer, status page, or support path. The win margin is 2.5 percent in the guides and 3 percent in the site's llms.txt, a file written for AI agents. Registration cost and miner hardware are not stated.
Go deeper
Sources for this explainer
Metaphor: a playground game of king of the hill. 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
- Albedo
- github_repo
- github.com
- subnet_contact
- arbos@bittensor.com
- subnet_url
- us-east-1.hippius.com
- discord
- @arbos
- description
- Alchemical intelligence
- additional
- not set
Stakers and validators●●●●● 2.3
Should I allocate here? · rank 27 of 32 for this audience
Q1What it is●●●●● 3
King-of-the-hill trajectory-distillation subnet on Bittensor (netuid 97, finney).
Output is a crowned fine-tuned Qwen3.6-35B-A3B checkpoint, no unit or price, stated in a machine facing llms.txt linked from the front page nav, chain description reads Alchemical intelligence
Q2Who it is for●●●●● 1
nothing on users, customers, revenue or usage in own materials, the only sign of use is a Chat with the King link that opens an Open WebUI instance, third-party pages add nothing on demand
Q3How it resists gaming or fails●●●●● 3
SCORING.md names anti-gaming steps (reference leak filtering, prompt injection stripping, looped trajectories scored zero) and a failure rule (eval fails below 0.8 valid samples), lives in the repo docs folder with no README pointing to it, needs the mechanism vocabulary, llms.txt says a 3 percent margin while SCORING.md and MINING.md say 2.5
Q4Identity and documentation●●●●● 2
identity 6 of 7 but the discord field is the handle @arbos not a link, subnet_url is a hippius redirect page to albedo.tech, github resolves to a repo with no README, no staker or validator page
Miners●●●●● 2.8
Can I compete, and what wins? · rank 27 of 32 for this audience
Q1What it is●●●●● 3
Albedo is a king-of-the-hill subnet for Qwen3.6-35B-A3B language models.
Miner fine-tunes that model, uploads to HuggingFace or Hippius and commits a v7 reveal on chain, model constraints stated (F16 or BF16, no py files, metadata byte for byte), no worked example, three clicks from the front page via a repo with no README
Q2Who it is for●●●●● 2
Registration costs recycled TAO
with no figure (chain burn 0.21 TAO today), no hardware statement for miners anywhere, competitiveness stated only as beating the king by the 2.5 percent margin in two separate evals
Q3How it resists gaming or fails●●●●● 3
duel scoring by checklist with four equal weighted tag categories, a 2.5 point absolute margin and two SUCCEEDED runs to promote, rolling 5 slot king chain for emissions in llms.txt, no example numbers, no immunity or deregistration policy, margin contradicts the 3 percent in llms.txt
Q4Identity and documentation●●●●● 3
step by step guide with albedo CLI commands (validate, upload, check, register, commit), repo pushed 2026-09-09, no releases, no dates on the guide, no README to reach it from
Buyers and enterprises●●●●● 1.3
Can I use this today? · rank 28 of 32 for this audience
Q1What it is●●●●● 3
Chat with the King
link to chat.albedo.tech which loads an Open WebUI app, no product page, no API, no example, the rest of the front page renders as loading placeholders
Q2Who it is for●●●●● 0
no pricing and no customer type anywhere in own or third-party materials
Q3How it resists gaming or fails●●●●● 0
no status page, SLA, support address or support path on the site or in the repo, the eval-service-status.md file is an internal runbook, the discord field is a handle not a link
Q4Identity and documentation●●●●● 2
no API reference, subnet_contact arbos@bittensor.com is set on chain but appears on no page fetched
Newcomers●●●●● 1.8
What is this and why does it matter? · rank 31 of 32 for this audience
Q1What it is●●●●● 2
Trajectory-distillation king-of-the-hill Bittensor subnet
is the repo blurb and the only prose a reader reaches from the front page, chain description Alchemical intelligence is a tag, no plain paragraph on the site
Q2Who it is for●●●●● 1
own materials contrast the duel with a benchmark run, not with a centralized service, third-party search snippets supply the only plain claim about permissionless competition
Q3How it resists gaming or fails●●●●● 2
sections named benchmarks, reign, queue, history and fails exist on the front page but render as loading placeholders to a fetcher, the chat app loads, no dated benchmark, no explanation of the numbers
Q4Identity and documentation●●●●● 2
subnet_name Albedo matches the site header ALBEDO, description is not a sentence, url resolves through a redirect page, no about or learn page, discord field is not a link
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.