Agentica RL Line (DeepScaleR / DeepCoder / DeepSWE)

The Sky Computing Lab's Agentica team scaled pure RL down and up: DeepScaleR-1.5B (Feb 2025) surpassed o1-preview on AIME via iterative-context GRPO; DeepCoder-14B (Apr 2025) hit o3-mini-level coding; DeepSWE (Jul 2025, Qwen3-32B) reached then-SOTA open-weight 59% on SWE-bench Verified with test-time scaling. Bases: R1-Distill-Qwen for DeepScaleR/DeepCoder, Qwen3-32B for DeepSWE. Fully open recipes and datasets via the rLLM framework; joint with Together AI (compute/marketing) — partial company co-attribution.