I RL-trained Qwen3.6-35B-A3B to RL-train small task-specific Qwen models. Fully open source! 🤓

👋 Training my first RL model last year was super fun, now I've RL-trained a model that RL-trains other models... wild times! The agent gets a task, writes the full training job (environment, reward, dataset, hyperparameters), and submits it to real GPUs. When the model it trained scores higher on a hidden eval, the agent gets rewarded. An RL loop with RL loops inside it! 🤯 What I did: Built a harness where the trainer agent ( Qwen3.6-35B-A3B ) writes a complete prime-rl training job: a verifiers environment