Pluralis Research runs RL post-training on 14 consumer Macs

Pluralis Research has successfully executed the first RL post-training run where the entire rollout fleet operated on consumer Macs connected via the open internet. The setup utilized 14 Macs across four countries for int8 inference with MLX, while a single B200 handled bf16 gradient updates.

• Rollouts synchronized through Cloudflare R2 over ordinary home internet without datacenter interconnects.
• PULSE transferred int8 weight deltas (approx. 82 MB) instead of full checkpoints to manage the off-policy gap.
• A DPPO-style probability gate filtered out tokens with significant probability drift between rollout and trainer models.
• On PaperSearchQA, cover pass@1 increased from 29% to 63%, and search rate rose from 22% to 84%.

This approach demonstrates that training on distributed consumer hardware is viable, potentially keeping open models accessible as frontier models move behind closed APIs.