Phonon-2 is an English speech recognition model from Fermion Research, a quantized derivative of NVIDIA's parakeet-tdt-0.6b-v3 distributed in MLX format. The card describes it as the most accurate open English speech recognition model under 900 MB. The download is 164 MB, against 2,508 MB for the full-precision Parakeet TDT 0.6B v3 teacher, and the encoder holds each weight at one of five learned levels in about 2.1 bits. Across the seven English sets of the Open ASR Leaderboard it averages 5.21% word error, compared with 4.96% for the teacher, 5.69% for Parakeet Redux (178 MB), 6.58% for Whisper large-v3-turbo and 6.56% for its predecessor Phonon-1. The teacher is ahead on most individual sets, while Phonon-2 scores 2.46 on VoxPopuli against the teacher's 3.19 and 9.37 on AMI against 9.42. The teacher and Whisper rows are the leaderboard's published results, and the other rows were run with its code on the full test sets. Speed figures from the card: an hour of audio in about 20 seconds on an M5 MacBook Air (174x realtime), 143x on eight Zen 5 cores, and 6,680x on one H100 in batches of 128. The fermion-research package provides the phonon command for file transcription with optional word timestamps (--json) and hotword biasing, live microphone transcription on Apple silicon, and an OpenAI-compatible HTTP server. It runs on Apple silicon through MLX, on Linux (x86-64 and Arm), Windows and Intel Mac CPUs, and in CPU and NVIDIA CUDA Docker images. The model is also the engine inside Fermion's Detta dictation app for Mac. Weights are CC-BY-4.0, matching the original, and the command line and engine code are Apache-2.0. The tokenizer and output conventions (punctuation, casing, numerals) are inherited from the original, and the card lists English as its only language.