# HiveCoder3.2 `theailearner/HiveCoder3.1` fine-tuned on its own successful AgenticBees runs (rejection sampling): tasks with automatic checks, attempts kept only if the check passed and the run was clean. Serve with vLLM: `--tool-call-parser qwen3_coder --reasoning-parser qwen3`. Held-out tasks, before -> after: solved 39% -> 64%, junk 0% -> 0%, invented tools 0% -> 0%. Task data includes MBPP (google-research-datasets/mbpp, CC-BY-4.0).