A community-owned alternative to OpenAI and Anthropic, built on Bittensor.
Base is a Bittensor subnet building frontier AI in the open. Where OpenAI and Anthropic build behind closed doors, Base turns model research, data collection, and training into a set of open, competitive challenges run across a decentralized network of miners.
Every hard problem in the AI pipeline, from discovering better architectures to gathering the freshest training data, is framed as a challenge. Miners around the world compete to solve it, contribute real, working code, and are incentivized on-chain for the quality of their work. Validators score contributions and reward the best, so the network improves continuously without a single company owning the outcome.
The result is a lab without walls: the same ambition as the big closed labs, but transparent, permissionless, and owned by its contributors.
The core difference: closed labs are limited to a handful of salaried employees competing internally, with no external pressure and no merit-based upside. Base opens the same research to the entire world and rewards every contribution on-chain, so the best ideas can come from anyone, anywhere.
| Closed labs (OpenAI, Anthropic) | Base | |
|---|---|---|
| Who competes | Only internal employees | The whole world, permissionlessly |
| Talent pool | A few hundred hires | Thousands of miners, globally |
| Incentives | Fixed salaries, no merit reward | On-chain rewards for the best work |
| Ownership | Corporate, centralized | Community-owned |
| Data | Proprietary pipelines | Decentralized, continuously crawled |
| Transparency | Black box | Every contribution is public code |
Base is not one model. It is the machinery that produces models, opened up to everyone.
Base coordinates the full AI research loop through specialized challenges. Each one targets a distinct part of building better models:
Decentralized neural architecture search. Miners submit architectures and training recipes to discover scalable AI improvements, evaluated competitively so the strongest ideas rise to the top.
The research engine of Base: how we find better ways to build models.
Miners compete to crawl the web on demand, following links and extracting page content, incentivized to relay accurate, up-to-date data whenever it's requested.
The freshness engine: recent, real-world data flowing straight into training.
Developers are incentivized to create diverse, high-performance datasets, evaluated in isolated environments and rewarded on quality and utility.
The refinement engine: turning raw crawled data into clean fuel for training.
A platform challenge where developers run and monetize terminal-based AI agents, evaluated in isolated environments and rewarded through competitive performance.
Community-driven bug discovery and software improvement, with rewards based on impact and quality.
- A challenge is issued - research, crawl, dataset, or agent task.
- Miners compete - they contribute real code, crawled data, or model improvements.
- Validators score - quality is measured in isolated, verifiable environments.
- The network rewards - the best work earns on-chain incentives, and the whole system gets smarter.
Prism discovers how to build better models, Relay crawls the web for the most recent data, Data Fabrication turns it into training-ready datasets, and the network trains and improves, in the open.
Research, data, and training, decentralized and incentivized, competing head-to-head with the closed labs.
Documentation · Core Subnet · Prism · Relay
Frontier AI shouldn't be owned by a few. Base is building it for everyone.