Continued pretraining ("midtrain") of zai-org/GLM-4.5-Air-Base on a 50/50 mixture of SWE-agent-trajectory tokens (SWE-ZERO OSS trajectories + own distillation) and StarCoder-v2-anchored code / math / STEM / web text (dolmino mix, nemotron-cc-math). 184 train datasets; 16k sequence length; trained for 19,160 steps.

This is the initialization checkpoint used by the downstream SFT-v3 -> RL pipeline (the "bigrun"). At matched SFT steps, initializing from this checkpoint instead of the raw base improved SWE-bench-Verified by +6 to +9 points.