The Open-Source AI Stack
RSS

The Stack · Infrastructure · Open source

Prime Intellect

Distributed training framework and compute aggregation (INTELLECT-1, INTELLECT-2).

Apache 2.0 · research · Project site → · GitHub →

Prime Intellect is a startup focused on decentralized training of frontier-scale AI models. They aggregate compute across providers (their own and partners) and ship open frameworks for bandwidth-efficient distributed training over the public internet. Their `prime` repo packages the training stack. Prime Intellect matters because they are the leading datapoint on whether you can train a competitive open-weights model without a hyperscaler-class cluster. INTELLECT-1 (released November 2024) was the first reproducible decentralized pretrain of a 10B+ model across continents. INTELLECT-2 (mid-2025) followed at 32B with reinforcement-learning post-training distributed globally. Compared to other distributed-training bets: Nous Research (DisTrO optimizer, similar bandwidth- efficient theme), Templar (Bittensor subnet that ran a 72B distributed pretrain), Pluralis and Gensyn (earlier-stage with different technical approaches). Production-readiness is research-to-pilot. The technique works and ships models; the honest ceiling is roughly an order of magnitude behind centralized frontier training on iso-cost, iso-time. Prime Intellect's strategic angle is "make distributed training real, then push toward parity." Funded by venture capital plus partnerships; the open framework is independent of the company's commercial roadmap.

Sources

Want a follow-up? Ask the chat about Prime Intellect in context. It will compare to siblings at the same layer and ground every claim in the wiki.

Other projects at the Infrastructure layer

13 siblings · ordered open first