k3 RL Run Reportedly Used ~50M Sandboxes With Millions Concurrent, Checkpoints Merged Across Scaffolds

stochasticchasm · x · 2026-09-11

Based on the k3 report, stochasticchasm notes the RL run used roughly 50 million sandboxes over its full duration, with millions running concurrently. He adds this is a second data point (after MAI-thinking-1) of models continuing RL beyond collapse, and flags the aggressive practice of merging checkpoints across different scaffolds/harnesses.

Related event: DeepSeek V4 training details: 10T+ tokens, 1M context, stable RL(3 posts)→

Original post →

More from Infra

Infra channel →