EvolAI
SN47AI evolution network exploring automated model improvement
Miners train language models and publish them to HuggingFace. Validators score them on-chain. Best model wins the rewards.
// Decentralizing the LLM benchmark.
EvolAI (SN47) is a Bittensor subnet for competitive language model training. Miners train their own LLMs and publish them publicly on HuggingFace. Validators then evaluate those models and assign scores that determine reward distribution.
The simple version: It's like Kaggle for language models, but on a blockchain. Train a model, post it publicly, get scored by the network, earn rewards based on how your model performs.
Centralized equivalent: The Hugging Face Open LLM Leaderboard, or competitive benchmarking platforms like Kaggle.
How it works:
- Miners train language models using either a Transformer or Mamba2 architecture, publish them publicly on HuggingFace with "evolai" in the name, and register them on the subnet with a wallet and hotkey. Miners re-register each time they publish an updated model.
- Validators evaluate submitted models. The validator software requires a GPU with 80GB VRAM to run evaluations.
- The problem it solves: Centralized LLM leaderboards are run by single teams. That creates a trust gap: who picks the benchmark tasks, who can game the eval? EvolAI moves that scoring process onto an incentivized validator network.
- The opportunity: Open model evaluation is still a largely unsolved problem. A neutral, economically-incentivized benchmark could matter as the open-source LLM ecosystem scales.
- The Bittensor advantage: Validators are staked participants with skin in the game. Dishonest scoring costs them. Miner rewards track directly to model quality, which creates continuous competitive pressure to improve.
- Traction signals: The subnet is early. GitHub shows 31 commits from a single contributor, with the most recent commit just two days ago. No official website, Discord, or Twitter has been published on-chain. Price has moved sharply positive, up 31% over 7 days and 29% over 30 days, with net TAO inflows over the past week positive at around 160 TAO.
Category: Model Fine-Tuning | Centralized Competitor: Hugging Face Open LLM Leaderboard, Kaggle
The case for on-chain model evaluation is straightforward: leaderboards today require trusting the operator. EvolAI delegates that trust to an economically-incentivized validator network on Bittensor, where validators are financially penalized for inconsistency and rewarded for accurate scoring.
Mechanism:
According to the subnet's public repository, EvolAI runs two parallel tracks: transformer and mamba2. Miners choose a track, train a model, and upload it to HuggingFace as a public checkpoint. The model name must contain "evolai" to be eligible. Once uploaded, miners register their model on the subnet by specifying their wallet, hotkey, and track. They then receive a challenge from a validator to complete registration. After each new model version is published, miners re-register. Validators fetch the submitted models and run evaluations, requiring 80GB of GPU VRAM to do so.
The subnet's name and framing, "evolving AI systems," points to a design intent around iterative improvement: miners are pushed by the reward structure to continuously update and improve their models rather than coasting on a one-time submission.
EvolAI is thin and early. The root-side pool holds around 1,189 TAO, with a root proportion of 50.5%, meaning roughly half the pool depth comes from protocol subsidy rather than organic demand. Price discovery is still in progress. The subnet holds 0.63% of total network emissions. Daily buy volume has been around 421 TAO against sell volume of ~365 TAO, giving a positive net on 24-hour flow. Chain buys sit at 0.47% of emissions, a separate metric from the emission share itself and not a risk indicator at this level.
The single GitHub contributor and 31-commit codebase reflect very early stage development. The last commit was two days ago, suggesting active work. No website, Discord, or Twitter handle appeared in the on-chain identity, which limits community discovery at this point.
- Thin liquidity: The root-side pool holds around 1,189 TAO. Slippage on entry or exit will be material, and a single large position can move the alpha price significantly.
- Development concentration: One contributor on GitHub means key-person risk. If the developer steps back, development stalls with no visible backup.
- No public community presence: No official website, Discord, or social accounts have been registered on-chain. Discovery is limited to on-chain data and secondary aggregators.
- Early pool, high root proportion: Root prop at 50.5% means the pool is still in early price discovery and has not yet established deep organic demand. Price action here can be volatile relative to larger subnets.
- Undocumented evaluation methodology: The public repository describes how to submit and register models, but does not detail the specific evaluation criteria or benchmark tasks validators use. This opacity is a risk for miners trying to optimize their training and for anyone assessing the system's integrity.