# Into: Albedo A Bittensor subnet that runs distillation as a permanent tournament: upload a small coding model, duel the reigning champion on real software-engineering tasks, and take the crown only if a panel of AI judges says you beat it. // Dethrone the king or score nothing --- > New to Bittensor? Start here. Experienced users can skip to the full analysis. ### What is Albedo? Albedo (subnet 97) is a competition for shrinking large AI models into small, efficient coding models. Instead of one team training a model in private, anyone can submit a contender that tries to beat the current best model on a fixed set of programming problems. A new champion is crowned only when a challenger clearly outscores the old one. **The simple version:** It is a king-of-the-hill ladder for code-writing AI. There is always one reigning "king" model. You train a challenger, it duels the king on the same coding tasks, and if a panel of AI judges scores you higher by a clear margin, you become the new king. **Centralized equivalent:** No clean equivalent. The closest analogy is the model-distillation pipeline an AI lab runs internally to compress a big model into a cheaper one, except here it runs continuously, in the open, with a public scoreboard and the duel records published for anyone to learn from. **How it works:** - **Miners** train a small "challenger" model, upload it to decentralized storage, and register it on-chain to challenge the current king. - **Validators** stage head-to-head duels between the king and each challenger on a pinned set of coding trajectories, have a panel of language-model judges score the results, and crown a new king only when a challenger clears the win margin. --- ### Why This Matters - **The problem it solves:** Turning a big, capable model into a small, cheap one that codes nearly as well is valuable, but the work usually happens privately inside labs. Albedo turns it into an open, always-running contest where every duel is scored and the winning traces are published for the whole field to distill from. - **The opportunity:** Small coding models that punch above their parameter count are useful to a lot of people. A permanent open leaderboard, with the head-to-head records out in the open, is a public good for anyone working on model distillation. - **The Bittensor advantage:** The subnet pays out by measured result, not by promise. A challenger only earns by actually beating the king on identical tasks under independent judging, and the published duel traces compound: each round gives the next wave of miners better training data. - **Traction signals:** Early. The current codebase was rebuilt in late May 2026 and the competition is still spinning up, with little public discussion so far. The signal worth watching is the GitHub repository, where commits have been frequent through early June. --- ## Full Analysis **Category:** Model Fine-Tuning | **Centralized Competitor:** Internal lab distillation pipelines; an open Kaggle-style leaderboard for coding models Distilling a strong "teacher" model into a smaller "student" that keeps most of the capability is a standard technique, but it is normally a closed, in-house process. Albedo's bet is that an incentivized, continuously-running tournament produces a better stream of small coding models than any single team would, and that publishing every duel makes the resulting training data a shared resource. **Mechanism:** The design is king-of-the-hill. At any moment one model holds the crown. Miners build a challenger, typically by fine-tuning on the current king plus the published traces of past duels, then upload it to Hippius (the decentralized storage layer the subnet runs on) and post an on-chain reveal pointing at it. The current competition is locked to the Qwen3-4B size class: the README spells out that a challenger has to match the king's architecture exactly, down to hidden size, layer count, attention heads, and vocabulary, so duels compare training quality rather than model size. Validators run two pieces: a chain-and-orchestration process and a separate GPU evaluation server built on vLLM. The eval server pits the king and each challenger against the same pinned SWE-ZERO coding trajectories, and an ensemble of language-model judges, served through Chutes, scores the head-to-head results on a 0 to 100 scale. A challenger only takes the crown if its aggregate score beats the king by at least the configured win margin. Validators then set on-chain weights according to who holds the crown. To keep the contest honest, the pipeline screens submissions with a weight-fingerprint deduplication check, to catch copied models, and a prompt-injection probe before any duel runs. The "trajectory-distillation" half of the name comes from what happens after each duel: the full duel traces are published openly, so the winning side's completions become training data that anyone can distill from, no wallet required. Each round is meant to feed the next. The codebase, declared on-chain by the slot owner under the handle "arbos", is public under an MIT license and shows three contributors with steady commit activity through early June 2026. It was created in late May, which lines up with the competition still being in an early phase: at snapshot the subnet was drawing a small share of network emissions, around 0.6 percent by the smoothed Taoflow measure that Bittensor uses to size emissions from net staking flows. Albedo's alpha token traded near 0.02158 TAO with a pool holding roughly 6,084 TAO in depth. The subnet registered in March 2026 and sits inside its initial immunity window. --- ### Risk Factors - **Early execution risk:** The current mechanism is new. The repository was rebuilt in late May 2026 and the king-of-the-hill competition is only just getting going, so the design is unproven at scale and the early field of challengers is thin. - **Market and emissions:** Net TAO flows into the subnet were negative over the snapshot week and the alpha price has trended down over the past month. Under Bittensor's current Taoflow model, sustained net outflows pull a subnet's emission share lower, which a young subnet needs to reverse to keep funding participants. - **Deregistration, on the horizon:** Albedo registered in March 2026 and is still inside its four-month immunity window, so it cannot be deregistered yet. Once immunity lapses around mid-2026, a subnet sitting at a low EMA price among non-immune subnets becomes a candidate for automatic deregistration, so emission share will matter more then. - **External dependencies:** Judging runs through Chutes and model storage runs on Hippius. Both are central to the mechanism, so an outage or change at either service would directly affect the subnet's ability to run duels. --- Into the next one.