TAO's DeepSeek Moment: The Ascendancy of Templar (SN3)
Source: http://x.com/i/article/2033122059040206848
On February 3, 2025, a model named DeepSeek-R1 quietly went live on China
In the following month, due to its performance directly competing with top-tier closed-source models and training costs extremely low prices, it directly swept across the globe.
Triggered a sharp decline in US AI stocks and initiated the "DeepSeek" moment for Chinese AI.
On March 10, 2026, Bittensor's Subnet 3 Templar announced the completion of the largest LLM pre-training run in history - Covenant-72B
This is the largest decentralized LLM pre-training run in history:
7.2 billion parameters, on a dataset of approximately 1.1 trillion tokens, fully implemented through the network of Bittensor Subnet 3, permissionless, with over 70 independent nodes freely participating.
Bittensor has its own DeepSeek moment.
1. Templar (SN3): Paradigm Transformation from Data Acquisition to Core Training
Templar, formerly known as SN3 operated by Omega Labs, initially focused on the collection and mining of MultiModal Machine Learning data. With the evolution of the Bittensor mechanism, this subnet has completed a strategic leap from a "data porter" to a "model caster".
Currently, Templar is positioned as a global distributed large model pre-training infrastructure. It aggregates global heterogeneous computing power through the Incentive Mechanism, aiming to address the extremely high computational costs and centralized censorship issues in large model training. The successful delivery of Covenant-72B validates the maturity of this decentralized production model.
2. Covenant-72B: Breaking the Scale Ceiling of Decentralized Training
Covenant-72B is a milestone achievement produced by Templar, and also the largest dense architecture pre-trained model in the current decentralized network.
- Core Parameters:
It has 72 billion parameters and is pre-trained on a high-performance DCLM corpus.
- Performance Benchmarking:
In the evaluation of the base model, its performance is basically on par with Meta's Llama-2-70B.
- Instruction Optimization:
After fine-tuning, Covenant-72B-Chat demonstrates strong competitiveness in the IFEval (Instruction Following) and MATH (Mathematical Reasoning) dimensions, even outperforming closed-source models of the same scale on specific metrics.
- Inference Efficiency:
The model achieves an extremely high throughput of 450 tokens/sec, addressing the pain point of response latency in the practical application of large models.
3. SparseLoCo Algorithm: The Underlying Engine for Decentralized Training
When training a 72B-scale model in a common Internet environment, the biggest challenge lies in the communication bandwidth bottleneck between nodes. Templar has achieved a qualitative breakthrough by adopting the core algorithm SparseLoCo :
Extreme Compression:
The algorithm only selects 1%-3% of the core gradient components for transmission and quantizes the data to 2-bit, significantly reducing the demand for network bandwidth.
Low-frequency Synchronization:
Different from the step-by-step synchronization of traditional clusters, SparseLoCo allows nodes to perform 15 - 250 local iterations before global synchronization.
Error Compensation:
Through the local gradient accumulation mechanism, it ensures that the convergence accuracy of the model remains unaffected even when more than 97% of the information is lost.
This technological approach demonstrates that even without expensive dedicated line clusters such as InfiniBand, relying on globally distributed ordinary networks can also produce top-level intelligence.
4. Industry Evaluation and Market Reaction
Templar's technological achievements have attracted the attention of the mainstream AI community and Capital Markets:
- Industry Recognition:
Jack Clark, co-founder of Anthropic, classified Templar as the world's largest active decentralized training network in his analysis report and pointed out that its development speed has exceeded industry expectations.
Jason Calacanis (host of the All-In Podcast and well-known Silicon Valley investor) provided an in-depth introduction to Bittensor's mechanism in a recent blog post and hinted at everyone to buy it
- Institutional Layout:
Grayscale continues to increase its holdings of TAO and regards it as the core position in the decentralized AI track.
DCG founded Yuma, specifically focusing on accelerating the development of the Bittensor (TAO) ecosystem, which is regarded as DCG's largest and most direct bet on decentralized AI.
- Market Performance:
$TAO:Following Templar's announcement of completing the training of its 72B large model, TAO has since rallied by over 30%.
Templar has risen 75% within 7 days, becoming the top performer in capturing Bittensor's current Emission, with a current Market Cap of only 70m.
5. Subnet Investment Potential and Ecosystem Ceiling
Templar's success has opened up a brand new realm of imagination for the Bittensor ecosystem:
Breaking the Value Ceiling:
For a long time, the outside world has questioned that Bittensor is merely "air incentives." Templar has proven that the protocol can produce monetization-competitive productivity tools, shifting the valuation logic of TAO from "narrative-driven" to "product-driven."
Potential of Heterogeneous Computing Power:
With the development of "Heterogeneous SparseLoCo", consumer-grade graphics cards (such as RTX 4090) will be able to directly participate in the training of trillion-parameter models in the future, achieving equalization of computing power resources.
Deterministic opportunities for subnets:
Under the dTAO mechanism, subnets like Templar, which have hardcore technological barriers and can continuously produce high-performance models, have tokens with extremely high long-term allocation value.
Templar has marketcap of 70m, with FDV of 350m
Currently, the mainstream large model companies OpenAI is valued at 840 billion, Anthropic at 350 billion, and Minimax at 45 billion.
It's not that Templar can directly compete with these companies, but in the current circle where narratives are scarce, attention is dissipating, and people no longer believe in decentralization, the emergence of Templar undoubtedly injects a shot in the arm for decentralized AI.
Conclusion
Templar has demonstrated that decentralized environments can not only store data but also generate intelligence. Covenant-72B is merely a starting point. With the vertical integration of SN3 (pre-training), SN39 (computing power), and SN81 (reinforcement learning), the prototype of a decentralized OpenAI running on the blockchain has already emerged.
Since its inception to the present, the Crypto industry has disproven numerous narratives. Even once-popular decentralized storage, decentralized computing power, and decentralized computers all seem to have been disproven. However, it is gratifying that there are still projects firmly advancing on the path of decentralization and achieving results.
Templar's success is not only Bittensor's DeepSeek moment, but may also be Crypto's DeepSeek moment