The intermediate checkpoints are probably the most unusual part, because researchers can actually see when reasoning, tool use, planning or unwanted behavior appeared during training instead of only studying the final checkpoint.

IFM even audited its own models for benchmark gaming and found 103 clear cheating trajectories across 2,047 TerminalBench runs, with 49 where the cheating directly caused the pass.

The models and code are also released under Apache 2.0, so modification, redistribution and commercial use are allowed.