Mega-ASR is an open-source speech recognition model published under the xzf-thu GitHub account, first-authored by Zhifei Xie, built as a LoRA fine-tune on Alibaba's Qwen3-ASR-1.7B foundation model rather than a model trained from scratch. Its focus is robustness to messy real-world audio: the authors describe training across 7 atomic acoustic conditions (noise, far-field speech, obstruction, echo and reverberation, recording artifacts, electronic distortion, and transmission dropout) combined into 54 compound acoustic scenarios, using 2.4M training samples. The recipe combines two stages: A2S-SFT, a curriculum-style acoustic-to-semantic supervised fine-tuning process that progressively trains the encoder and aligner before fine-tuning the LLM and then the full stack jointly; and DG-WGPO, a WER-gated reinforcement-learning stage the README says trained the published weights, combining a WER accuracy reward with an anti-repetition gate and a dual-granularity reward. Only the A2S-SFT code ships in the repo today; DG-WGPO's training code is listed as a coming item, not yet published. At inference, a router decides per sample whether to mount the LoRA weights, since heavy exposure to high-WER data slightly degrades ordinary transcription. The repo ships offline Transformers inference, a WebUI slated for further optimization, a faster vLLM backend, and a vLLM streaming entrypoint for long-form audio. An evaluation script computes WER or CER over JSONL test sets, with an optional no-routing mode. The team also released the Voices-in-the-Wild-2M dataset and Voices-in-the-Wild-Bench benchmark, plus an arXiv technical report. In one README case study, Mega-ASR scored WER 47.1 on a degraded sample versus 100.0 for Qwen3-ASR (empty output), 86.1 for Gemini-3-Pro, 85.3 for Seed-ASR, and 92.5 for Whisper; the authors separately report up to nearly 30 percent gains over leading open- and closed-source models on their own compound benchmark, a claim from the paper rather than an independently verified figure. Licensing is unresolved: the Hugging Face weights repo is tagged Apache-2.0, and the README states the project will be released under Apache-2.0, but the repository ships no LICENSE file.