what's Teutonic (Bittensor Subnet 3)?
Teutonic is a specialized subnet on the Bittensor network designed for decentralized training of large-scale AI models, particularly large language models (LLMs). It builds on the foundational codebase of the former Templar subnet, incorporating advanced techniques such as heavy gradient compression to achieve near-centralized performance across distributed nodes. Miners compete in model updates through a new evaluation mechanism that supports heterogeneous hardware, enabling efficient scaling without uniform infrastructure requirements. The subnet's roadmap targets ambitious parameter scales, starting with a 1 billion parameter model in live training, progressing to 30 billion and 72 billion models in accelerated timelines, and ultimately aiming for a 1 trillion parameter LLM. This leverages Bittensor's incentive structure where validators and miners are rewarded in TAO based on contributions to model quality and training efficacy, fostering a competitive ecosystem for open-source AI development.
Teutonic-I launched, targeting a 1T parameter model; BIT-0011 introduces locked stake for ownership.
Launched Teutonic-I, a new version of Templar targeting development of a 1 trillion parameter model, with a 1 billion parameter model already tested.
BIT-0011 proposal introduces a locked stake and conviction mechanism where subnet owners lock alpha tokens for chosen periods, receive conviction scores updated every 30 days, and cannot unstake locked tokens while conviction exists, with the highest conviction score determining subnet ownership.