Zipcode
SN46AI-powered real estate data and intelligence, the oracle for property markets
The Bittensor subnet formerly known as RESI is now Zipcode on-chain, with a new consumer site at zipcode.ai. Same netuid (46), same owner, same price prediction mechanism, but the public-facing product is widening from a model competition into P2P real estate services and settlement.
// Open AVM competition, rebranded for the front door.
Zipcode (SN46) is a Bittensor subnet that incentivizes the development of accurate residential property price prediction models. Miners train machine learning models that estimate what US homes will sell for, and validators score those models against real sales data the models could not have seen at submission time. The team has rebranded the public-facing product from RESI to Zipcode while the on-chain mechanism stays the same.
The simple version: Imagine Zillow's Zestimate as an open competition. Anyone can submit a price prediction model, and every day the most accurate one against real recent sales takes almost the entire reward pool.
Centralized equivalent: Zillow Zestimate, Redfin Estimate, CoreLogic AVM, HouseCanary. Zipcode rebuilds the same automated valuation model surface as an open, swappable competition on public rails.
How it works:
- Miners train ONNX residential property price models, upload them to Hugging Face, and commit the model hash on-chain. Once committed, the model is locked.
- Validators download each committed model, run it against properties that listed and sold within the last 30 days, score by Mean Absolute Percentage Error (1 minus MAPE), and set weights so the earliest commit in the winner set takes 99% of emissions.
- The problem it solves: US residential real estate is priced largely by closed AVMs from a small number of vendors. Zipcode replaces that with a continuous open competition where the best public model wins, and grading uses sales that did not exist when the model was committed.
- The opportunity: Even small accuracy gains compound across mortgage underwriting, refinancing, home equity products, and property tax appeals. The team is positioning the subnet to feed a broader product surface, including P2P real estate services and settlement on zipcode.ai.
- The Bittensor advantage: Anti-overfitting is enforced by protocol. Models must be committed roughly 30 days before evaluation, so memorization of the test set does not pay. Copycats are also detected and zeroed out, and within the winner set the earliest committer wins, so cloning the leader cannot displace them.
- Traction signals: Active GitHub with 4 contributors and 67 commits, last push on 2026-05-21. Resi Labs signed a non-binding LOI with NextGen Digital Platforms in October 2025 for a 50/50 commercialization JV, which included a USD 100,000 alpha token purchase with a six-month lock-up.
Category: Other (Real Estate AI and Property Valuation) | Centralized Competitor: Zillow Zestimate, Redfin Estimate, CoreLogic AVM, HouseCanary
US residential real estate is roughly a 50 trillion dollar asset class, but the price models that move it (AVMs used by lenders, insurers, and tax assessors) come from a small number of vendors and are not open to inspection. Zipcode's bet is that an open competition with a hard temporal anti-overfitting rule can produce a better, more transparent AVM than any single closed vendor, and that the resulting model and data feed are valuable enough to anchor a broader real estate services brand.
Mechanism:
Miners train property price prediction models in any framework, export to ONNX (under 200MB, MIT-licensed via Hugging Face metadata), and commit the model hash to their hotkey on-chain. Once committed, the model is frozen. Validators run a daily evaluation at 18:00 UTC against properties that listed and sold within the last 30 days, run every eligible model in a sandboxed environment, and score by MAPE. A model with 8.5% average error scores 0.915.
Weight distribution uses a threshold-plus-commit-time rule. Validators identify the highest-scoring model, then a winner set of models within a configurable distance of that score. Within the winner set, the earliest on-chain commit takes 99% of emissions. Non-winners share 1% proportionally by score, and detected duplicates of an existing committed model receive nothing. The effect is that copying the current leader cannot dethrone them, and beating them by less than the threshold does not flip the king either.
Two transitions are running in parallel right now. The on-chain identity has been renamed from RESI to Zipcode, with the consumer site at zipcode.ai and positioning around P2P real estate services and settlement. The GitHub repo at github.com/resi-labs-ai/RESI-models is still RESI-branded but is actively developed: the last push was 2026-05-21, and recent commits include tie-breaking threshold updates, an exclusive-licensing change set, a 30-day model submission extension with optional image features, and a "100% burn" change. That last commit corresponds to the second transition: miner emissions are currently being fully burned, which is why TaoSwap reports zero active miners during this window. This is a deliberate switch by the team rather than abandonment, and prior commits show the burn lever has been moved before.
At 0.01036 TAO per alpha and roughly 13,968 TAO of root in the pool, SN46 sits in the smaller-cap tier of Bittensor. Market cap is about 51,967 TAO. The 30-day price is up roughly 41% and 90-day is up roughly 137%, against a 24-hour drawdown of 6.6%. Root proportion of 0.16 is low, meaning most of the pool depth is organic stake rather than protocol subsidy. Gini coefficient across the top 100 positions is 0.504 with an HHI of 0.0145, both consistent with relatively distributed ownership for a subnet this size. Chain buy rate sits at 1.70%, broadly in line with the emission EMA at 1.72%.
The roadmap, per the team's about page, targets a national property database, downstream AI products like seller intent prediction and market analysis, a third-party API ecosystem, and expansion into commercial and international markets. The NextGen LOI signed in October 2025 contemplates a 50/50 British Columbia JV to handle enterprise contracts, integrations, and revenue sharing, with a USD 100,000 alpha token purchase on a six-month lock-up as the initial step.
- Brand transition: On-chain identity, website, and product positioning have all moved from RESI to Zipcode, while the GitHub repo and parts of the documentation still carry the RESI name. Rebrands during a launch window add execution risk and can mute discovery while the new domain ranks.
- Burn window: The most recent commit moved miner emissions to 100% burn, which is why TaoSwap reports zero active miners. This is a deliberate lever rather than abandonment, but until it is reversed there is no real economic reward for new model submissions and the active miner count will sit at the floor.
- Execution: Zipcode is taking on incumbent AVM vendors with decades of data partnerships across lenders and assessors. The subnet has shipped a working competition and a public dashboard, but the consumer product on zipcode.ai and any downstream settlement product are net new and not yet in market.
- Concentration: Gini coefficient of 0.504 across the top 100 positions suggests moderately concentrated ownership or stake distribution. Large positions could significantly impact pool dynamics.
- Liquidity: Total 24-hour volume of about 3,312 TAO against a 52,000 TAO market cap means slippage on larger trades is real. The 7-day net flow of roughly +537 TAO is positive but modest, and emission share at about 1.04% is still small versus the broader subnet field.
Same subnet, new front door.
