Templar
Incentivized internet-wide LLM pre-training, no data center required
Templar trains large language models across contributors on commodity hardware.
The simplest explanation
Like a Wikipedia edit-a-thon for training AI: thousands chip in and build something massive.
Everything you need to understand Templar
No jargon. No whitepapers. Just the facts.
The Problem
Frontier model pre-training is locked behind massive centralized GPU clusters and closed labs; there is no permissionless way to pool global compute to train large models.
What Templar Does
Templar trains large language models across contributors on commodity hardware.
The Edge
Fully permissionless, whitelist-free distributed pre-training over commodity internet using SparseLoCo (sparsification + 2-bit quant + error feedback) for 146x comms reduction; produced an openly released 72B model (Apache license).
What's the closest mainstream equivalent?
The best way to understand a new technology is to compare it to something familiar.
Fully permissionless, whitelist-free distributed pre-training over commodity internet using SparseLoCo (sparsification + 2-bit quant + error feedback) for 146x comms reduction; produced an openly released 72B model (Apache license).
How does Bittensor make this possible?
Templaruses Bittensor's incentive layer to build something no single company could run alone.
Miners compete
Participants (called miners) on the Templar subnet compete to produce the best outputs for decentralized ai training tasks. Anyone with the right hardware can join.
Validators score the work
Validators continuously evaluate miner outputs against objective benchmarks. The best performers rise, the worst are replaced. There's no human committee — the protocol decides.
Rewards flow to the best
Miners are paid in the Templar alpha token in proportion to how good their work is. This creates a continuous competitive pressure that drives quality up and cost down — structurally, not just as a promise.
The whole network benefits
Because every participant is aligned toward the same goal — producing the best decentralized ai training results — the network improves continuously without requiring a central team to manage it.
Frequently Asked Questions about Templar
What is Templar on Bittensor?
Templar is Subnet 3 (SN3) on the Bittensor network — a decentralized AI protocol built on the TAO blockchain. Templar trains large language models across contributors on commodity hardware.
What problem does Templar solve?
Frontier model pre-training is locked behind massive centralized GPU clusters and closed labs; there is no permissionless way to pool global compute to train large models.
How is Templar different from A decentralized substitute for training frontier LLMs on a single centralized GPU supercluster (e.g. Meta/OpenAI training runs)?
Fully permissionless, whitelist-free distributed pre-training over commodity internet using SparseLoCo (sparsification + 2-bit quant + error feedback) for 146x comms reduction; produced an openly released 72B model (Apache license).
What is the Templar token?
The Templar subnet has its own alpha token on Bittensor's dTAO system. It trades in the Bittensor liquidity pool and its price reflects market demand for the subnet's services.
Is Templar a good investment?
AlphaGap tracks Templar using its aGap score — a composite of development activity, token flow, and social signals. This page is for informational purposes only and is not financial advice. Always do your own research before making any investment decisions.
Want real-time intelligence
on Templar?
AlphaGap tracks signals, whale flows, developer commits, and the aGap score for every Bittensor subnet — updated continuously. Find the alpha gap before everyone else.
Free preview · Pro from $29/mo · Premium from $49/mo
