Qwen 7B fine-tuned with GRPO on NalandaJEENEETBench. GPU inference runs on Modal; the first request after idle has a short cold-start delay.