Templar: first permissionless 1.2B LLM training run on Bittensor SN3
Quote271 words
Explains how Templar's Gauntlet trained a 1.2B LLM permissionlessly: miners train every 84 seconds, a two-phase evaluation scores loss reduction, and only the top 15 contributions per round are applied. Benchmarks include HellaSwag 51.0%, PIQA 71.4% and ARC-E 59.2%.
Key claims
A 1.2B-parameter LLM was trained entirely permissionlessly on subnet 3
Only the top 15 contributions per round are applied
HellaSwag 51.0%, PIQA 71.4%, ARC-E 59.2%
The world's first truly decentralized LLM training run just happened.
A 1.2 billion-parameter LLM was trained entirely permissionlessly, proving that artificial scarcity caused by spiraling LLM training costs can be overcome.
Powered by SN3 @tplr_ai on Bittensor.
TL;DR:
→ Templar's Gauntlet introduces the first fully permissionless, decentralized training environment for large language models, enabling anonymous global peer contributions without registration or centralized control.
→ Miners on subnet 3 collaboratively trained a 1.2B-parameter model by competitively reducing loss. Every 84 seconds, miners perform real training work on uniquely assigned data subsets, measuring the model’s performance ("loss") and calculating pseudo-gradients.
→ Gauntlet evaluates miner contributions using a two-phase system: Primary Evaluation assesses value through loss reduction and OpenSkill (Elo-like) rankings; Fast Evaluation ensures quick compliance and synchronization checks.
→ Only the top 15 highest-quality contributions per round are applied to the global model, creating intense competition among miners to optimize their training methods and computational efficiency.
→ Unique data subsets assigned per peer each training round prevent cheating, ensuring genuine computation and integrity through comparative loss scoring.
→ Implements robust Byzantine fault tolerance via gradient normalization and signed aggregation, safeguarding the training process against malicious or faulty inputs.
→ Utilizes straightforward and resilient cloud-based (S3) communication for global synchronization, timestamping, and transparent peer coordination.
→ Achieved competitive benchmarks (HellaSwag 51.0%, PIQA 71.4%, ARC-E 59.2%), outperforming comparable centralized models.
→ Economically incentivizes fewer, higher-quality contributors, directly aligning rewards with the impact of contributions on model performance.
→ Represents a critical evolution from centralized AI monopolies towards decentralized, incentive-driven collective intelligence in foundational AI development.
DYOR anon.
Not familiar with Bittensor yet?
Check here ↓
QUOTING @tplr_ai
The cost of training large language models has created an artificial scarcity. This isn't a technical limitation—it's a failure of coordination.
The cost of training large language models has created an artificial scarcity. This isn't a technical limitation—it's a failure of coordination.
1 image on X
How the covered assets moved · since 28 May 2025
Prices as of 23 Oct 2026This explainer walks through how things work without making a call on the token, so price moves since publication are shown as market context.
Covered in this piece
Since then
Zero has written 9 more pieces on Bittensor and $TAO since this one.
- Watchlist additions: TAO, PURR, EDEL and BPOn Bittensor
- I Built a Research Desk Staffed Entirely by AI Agents. It Works.Covers $TAO
- AI infra projects on watch: Bittensor, Sentient, NEAR, Ritual and moreOn Bittensor
- Why DeAI matters across the AI stack: seven layers from incentives to computeOn Bittensor
Daily closes in USD from DefiLlama and Yahoo Finance. Only a bullish or bearish thesis is read as a call; explainers, maps and mentions are context. Not financial advice.