Bittensor is the hope of the entire crypto community

Author: 0xai

Special thanks to @DistStateAndMe and their team for their contributions to open-source AI models, as well as for their valuable advice and support for this article.

Why you should pay attention to this report

If “decentralized AI training” has gone from impossible to possible, how underestimated is Bittensor?

In early 2026, the entire crypto community is filled with fatigue.

The afterglow of the last bull market has long faded, and talent is rapidly flowing into the AI industry. Those who once talked about “the next 100x” are now discussing Claude CodeOpenclaw. “Crypto is a waste of time”—you might have heard this more than once.

But on March 10, 2026, a subnet called Templar within Bittensor quietly announced something.

Over 70 independent participants from around the world, without central servers or big companies coordinating, relying solely on crypto incentives, collaboratively trained a large AI model with 72 billion parameters.

The model and related papers have been published on HuggingFace and arXiv, with data openly verifiable.

More importantly: in multiple key tests, this model outperformed Meta’s similarly sized models trained with heavy investment.

After the announcement, TAO’s price remained silent for nearly 2 days. It only started to surge on the third day, and after 6 days, it continued to rise, with a total increase of about +40%. Why was there a 2-day delay?

The core argument of this report is: crypto investors see “another open-source model” and think it’s not as good as GPT or Claude used daily; AI researchers don’t pay attention to crypto. The gap between these two communities is creating a cognitive arbitrage window.

Reading Framework

This report is divided into two logical parts:

Part I — Technological Breakthrough: Explains what SN3 Templar actually achieved and why this is significant in AI and crypto history.

Part II — Industry Significance: Explains why this event indicates that the Bittensor ecosystem is systematically underestimated and why Bittensor is the hope of the entire crypto community.

Part I: Breakthrough in Decentralized AI Training

1. What does SN3 do?

What is needed to train a large language model?

Traditional answer: Build a massive data center, buy tens of thousands of top GPUs, spend hundreds of millions of dollars, and have a team of engineers coordinate it all. This is the approach of Meta, Google, OpenAI.

SN3 Templar’s approach: Let scattered individuals around the world each contribute one or several GPU servers, piece together their computing power like a puzzle, and jointly train a complete large model.

But there’s a fundamental challenge: if participants are globally distributed, distrust each other, and network latency is unstable, how to ensure the training results are effective? How to prevent laziness or cheating? How to motivate continuous contribution?

Bittensor provides an answer: use TAO tokens as incentives. The more effective a participant’s gradient (which can be understood as “contribution to model improvement”), the more TAO they earn. The system automatically scores and settles rewards without any centralized organization.

This is SN3 (Subnet #3), code-named Templar.

If Bitcoin proved that decentralized “money” is possible, SN3 is proving that decentralized “AI training” is also possible.

2. What has SN3 achieved?

On March 10, 2026, SN3 Templar announced the completion of training a large language model called Covenant-72B.

What does “72B” mean? 720 billion parameters. Parameters are the “knowledge storage units” of AI models; the more, the smarter the model. GPT-3 has 175 billion, Meta’s open-source LLaMA-2 has 700 billion. Covenant-72B is in the same size range as LLaMA-2.

How large was the training scale? About 1.1 trillion tokens, roughly equivalent to 5.5 million books (assuming 200,000 words per book).

Who participated in the training? Over 70 independent miners contributed computing power (with about 20 nodes synchronized per round). Training started on September 12, 2025, and lasted about 6 months. No central server or unified organization coordinated it.

How did the model perform? Using mainstream AI benchmarks for comparison:

Data sources: HuggingFace 1Covenant/Covenant-72B-Chat model card

• MMLU (57 subjects, comprehensive knowledge): Covenant-72B 67.35% vs Meta LLaMA-2 63.08%
• GSM8K (math reasoning): Covenant-72B 63.91% vs Meta LLaMA-2 52.16%
• IFEval (instruction-following ability): Covenant-72B 64.70% vs Meta LLaMA-2 40.67%

Fully open-source under Apache 2.0 license. Anyone can download, use, and commercialize without restrictions.

Academic backing: Paper submitted [arXiv 2603.08163], core techniques (SparseLoCo optimizer and Gauntlet anti-cheating mechanism) presented at NeurIPS Optimization Workshop.

3. What does this achievement mean?

For the open-source AI community: Previously, due to funding and computing barriers, training 70B-level large models was the domain of a few big companies. Covenant-72B proves for the first time that a community can train models of this scale without any centralized funding support. This changes the boundaries of who is qualified to develop foundational AI models.

For AI power structures: The current landscape of foundational AI models is highly centralized—OpenAI, Google, Meta, and Anthropic control the most powerful models. Decentralized training’s success means this moat is no longer invulnerable. The premise that “only big companies can do foundational models” is being shaken for the first time.

For the crypto industry: This is the first time a crypto project has made a genuine technical contribution to AI, rather than just riding the hype. Covenant-72B comes with HuggingFace models, arXiv papers, and open benchmark data. It sets a precedent: crypto incentive mechanisms can serve as infrastructure for serious AI research.

For Bittensor itself: The success of SN3 transforms Bittensor from a “theoretically feasible decentralized AI protocol” into a “practically validated decentralized AI infrastructure.” This is a qualitative leap from 0 to 1.

4. The historical significance of SN3

Decentralized AI training is not new—SN3 is not the first to explore this path. But SN3 has reached places others have not.

Evolution of decentralized training:

2022 — Together GPT-JT (6B): Early exploration, proved multi-machine collaboration feasible

2023 — SWARM Intelligence (~1B): Proposed heterogeneous node collaboration framework

2024 — INTELLECT-1 (10B): Cross-institution decentralized training

2026 — Covenant-72B / SN3 (72B): First large model surpassing centralized training on mainstream benchmarks

In 4 years, parameters grew from 6B to 72B—a 12x increase. But more important than size is quality—earlier projects mainly demonstrated “it can run,” while Covenant-72B is the first decentralized large model to outperform centralized models on mainstream benchmarks.

Key technological breakthroughs:

99% compression ratio (>146x): When participants upload gradients, originally gigabyte-scale data, SparseLoCo compresses it over 146 times, akin to compressing an entire season of TV into a single image with minimal information loss.

Only 6% communication overhead: 100 participants working together spend only 6% of their time on “coordination,” with 94% dedicated to actual training. This solves one of the biggest bottlenecks in decentralized training.

5. Is decentralized training underestimated?

Let’s look at data before judging.

Evidence of underestimation:

• MMLU 67.35% vs LLaMA-2 63.08%
• MMLU-Pro 40.91% vs LLaMA-2 35.20%
• IFEval 64.70% vs LLaMA-2 40.67%

Models trained via decentralized methods surpass Meta’s heavy-investment LLaMA-2-70B.

Compared to current top open-source models (honestly facing the gap):

• MMLU: Covenant-72B 67.35% vs Qwen2.5-72B 86.8% vs LLaMA-3.1-70B 83.6%
• GSM8K: Covenant-72B 63.91% vs Qwen2.5-72B 95.8% vs LLaMA-3.1-70B 95.1%

The gap is about 20-30 percentage points.

But the trend is more important than snapshot data:

• 2022: Best decentralized model was 6B, no MMLU testing.
• 2026: 72B model, MMLU 67.35%, surpassing Meta’s models.

In 4 years, decentralized training has evolved from “concept experiment” to “performance comparable to centralized training.” The slope of this curve is more noteworthy than any single benchmark number.

Moreover, the gap in deep reasoning performance of Covenant-72B has planned solutions—SN81 Grail will handle post-training reinforcement learning (RLHF) to align and enhance capabilities. This is the key step that improved GPT-4 over GPT-3.

Heterogeneous SparseLoCo is the next milestone: currently, SN3 requires all miners to use the same GPU model. The upcoming Heterogeneous SparseLoCo will enable mixed hardware (B200, A100, consumer GPUs) to participate in the same training task. Once achieved, the compute pool for the next training round will expand significantly.

Decentralized training has crossed the feasibility threshold. The current benchmark gap is an engineering problem to optimize, not a fundamental theoretical barrier.

Part II: The market still doesn’t understand this

TAO Price Timeline

After SN3’s announcement, TAO’s price movement reveals this recognition lag:

Note the 2 days of silence (3/10 → 3/12): the announcement was made, but the price hardly moved.

Why the delay?

Crypto investors saw the news as “Bittensor SN3 completed training an AI model”—but they might not understand the technical significance of “72B decentralized training surpassing Meta on MMLU.”

AI researchers understand the technical significance but don’t care about crypto.

The recognition gap between the two communities creates a 2-3 day price lag window.

Most crypto investors’ understanding of Bittensor remains stuck in the previous cycle. Today, over 79 active subnets on Bittensor cover diverse fields like AI agents, computing power, AI training, AI trading, and robotics. When the market revalues Bittensor’s ecosystem breadth, this recognition gap will be corrected—and the correction often manifests as a sharp price surge.

Bittensor’s valuation dislocation

Placing Bittensor in the broader industry context:

SN3 has proven: Bittensor can accomplish decentralized large model training.

If future AI requires open, permissionless training networks, the only practically validated infrastructure candidate is Bittensor.

The market is pricing a foundational AI infrastructure network based on application-layer project valuation logic.

Even within crypto: Bitcoin’s market share in the entire crypto space has long been 50-60%, while Bittensor’s share in the crypto AI track is only about 11.5%.

As the market re-understands Bittensor’s role in AI infrastructure, this dislocation will inevitably be corrected.

Conclusion: Bittensor is the hope of the entire crypto community

If Covenant-72B proves one thing, it’s that:

Decentralized networks can coordinate not only capital but also compute and cutting-edge AI R&D.

In recent years, crypto’s role in AI narratives has been mostly marginal. Many projects rely on conceptual packaging, hype, or capital narratives but lack verifiable technical output. SN3 is a clear exception.

It didn’t introduce a new token narrative nor packaged an “AI + Web3” application layer product, but achieved something more fundamental and difficult:

Training a 72B-scale model without centralized coordination.

Participants come from all over the world, trust isn’t required among them; the system relies on on-chain incentives and verification mechanisms to automatically coordinate contributions and reward distribution.

Crypto mechanisms have, for the first time, organized real productivity in AI.

Many still don’t grasp SN3’s historical significance. Just as many didn’t realize that Bitcoin proved not just “better payments” but a trustless value consensus without central authority.

Today, many still see only benchmarks, model releases, or a price rally.

But the real change is that Bittensor is proving:

Crypto is not just about issuing assets; it can organize production.

Crypto is not just about trading attention; it can produce intelligence.

Open-source communities can contribute code, academia can contribute papers, but when it comes to large-scale training, long-term collaboration, cross-region scheduling, anti-cheating, and reward sharing, good intentions and reputation systems are far from enough:

Without economic incentives, there’s no stable supply.

Without verifiable rewards and punishments, there’s no long-term cooperation.

Without tokenized coordination mechanisms, a truly global, permissionless AI production network cannot form.

So, is Bittensor underestimated? The answer is not “possibly,” but “significantly and systematically underestimated.”

In the ongoing debate over “Does crypto still matter,” Bittensor is providing the most compelling answer to the entire industry.

And for this reason: Bittensor is the hope of the entire crypto community.