Covenant72B

Covenant72B is the largest permissionless collaboratively trained language model trained entirely from scratch at the 72 billion parameter scale.

It is being trained with 20+ globally distributed participants coordinated via decentralized infrastructure on the Bittensor blockchain.

The latest base model files are available on the base branch. Future checkpoints will be updated here.

---

Training Details

Property	Value
Model size	72B
Architecture	LLaMA-style
Target token budget	1.2T (420B for current checkpoint)
Compute participants	20+
Minimal compute per participant	8×B200 or equivalent
Dataset	DCLM-baseline
Optimizer	SparseLoCo (communication-efficient optimizer)

---

Performance on Benchmarks

All results are 0-shot acc-norm (%) unless noted.

Model	Compute Environment / Permissions	Size	Tokens	ARC-C	ARC-E	PIQA	OpenBookQA	HellaSwag	Winogrande (acc)	MMLU (acc)
Intellect-1	Internet / Whitelist	10B	1T	44.8	71.6	77.7	43.6	70.5	63.1	32.7
Psyche Consilience-7Y9	Internet / Whitelist	40B	1.2T	31.1	55.8	76.1	34.8	63.7	57.0	24.2
Covenant72B (Checkpoint-Two)	Internet / Permissionless	72B	420B	53.84	77.74	80.58	44.60	77.08	71.43	47.49
LLM360 K2 ckpt_108	Centralized Cluster	65B	420B	45.73	70.54	80.90	43.20	78.23	71.90	50.01
LLM360 K2 Stage 1	Centralized Cluster	65B	1.4T	53.84	75.93	82.48	48.00	82.81	76.64	63.90
LLaMA-2-7B	Centralized Cluster	7B	2T	45.90	74.58	75.92	44.20	75.92	68.90	40.86
LLaMA-2-70B	Centralized Cluster	70B	2T	57.59	80.77	82.92	48.60	83.86	77.58	65.56

---

For more details, refer to Checkpoint One on Templar Research.