Templar AI and Covenant-72B

Short answer: Templar was Bittensor subnet 3, a network where anyone could contribute GPUs to pre-train a language model. The team, now called Covenant AI, trained Covenant-72B that way, left Bittensor in April 2026, and runs its chat on its own GPU cloud, Basilica. Serving models across volunteers’ hardware is the piece it says it is working on now. Nothing about it has shipped yet.

What Templar trained

Covenant-72B is a 72.7-billion-parameter language model pre-trained on about 1.1 trillion tokens by more than 70 contributors who did not need anyone’s permission to join. Each contributor’s update was scored by how much it lowered the model’s loss before it was accepted, so a peer could not poison the run by sending junk. Updates were compressed hard enough to cross ordinary internet links. The paper is arXiv 2603.08163 (March 2026), and the model compares itself to LLaMA-2-70B.

Contributors needed datacenter hardware (the training repository lists eight B200s per miner), not home GPUs or browsers.

Later work from the same team:

• Crucible, a pre-training platform; its first production run trained an 8B model on 50.5 billion tokens across 48 A100s (September 2026).
• PULSE, a way to send only the changed weights to inference workers during reinforcement learning, now used in Hugging Face’s TRL and vLLM.

The Bittensor split

On 2026-04-10 the team announced it was leaving Bittensor, after a dispute with Bittensor’s co-founder over subnet emissions (statement). It had run three subnets: 3 (Templar, pre-training), 39 (Basilica, compute) and 81 (Grail, post-training). Covenant AI’s site still lists Templar, Basilica and Grail as its products.

Inference: what you can use today

• Covenant-72B chat runs at tplr.ai/chat on Basilica, which is a GPU cloud: the model runs on rented servers, not on a peer-to-peer network.
• Distributed serving: in August 2026 the team wrote that “serving open models on distributed hardware is the piece we are working on now.” No code, paper or date has been published.

If you want peer-served models now, Obscurify.ai’s P2P sharing serves open models up to 9B from volunteers’ browser tabs, and the tabs earn credit for any model. It is much smaller than a 72B network, and it works today.

How to use these models privately

Whatever network serves a model, the machine that runs it reads your prompt. What you can control is whether it knows who you are:

• Use a front end that doesn’t tie requests to an email or phone number. An Obscurify.ai account is a random 16-digit number.
• Pay in a way that doesn’t follow you around: Monero, Bitcoin or blind-signed vouchers.
• Connect over Tor if your network address matters.

Obscurify.ai tracks Covenant AI’s code and announcements and will add its distributed serving when it ships. The wider picture is on decentralized AI inference.

Common questions

Is Templar still a Bittensor subnet?

Not under the team that built it. Covenant AI left Bittensor in April 2026. Reports say subnet 3 continues under different operators.

Can I run Covenant-72B myself?

It is a 72B model, so it needs roughly 40 GB of GPU memory even at 4-bit precision. That rules out a single consumer GPU or a browser tab.

Can I contribute my GPU to Templar?

Its training runs needed datacenter GPUs, and its training repository is marked dormant. To put a home GPU to work on AI today, P2P sharing runs in a browser tab.

Does Templar have a token?

Contributors were paid in Bittensor’s TAO and subnet tokens while it ran on Bittensor. Basilica’s documentation still mentions funding credits with TAO. Covenant AI has not published how payment will work for distributed serving.

Last updated 2026-11-06. Plan prices and model lists on this page are read live from the site.