Qwen3.5-0.8B-parakeet-52k-FullFT

Quality champion — best semantic match. BF16 full weights. Convert to MLX with mlx_lm convert --hf-path . -q --q-bits 8.

Training details

Base model	Qwen/Qwen3.5-0.8B
Recipe	Full fine-tune (BF16)
Training data	rdsm/parakeet-stt-redone
Format	BF16 safetensors

Benchmark (100 held-out pairs, Codex judge vs GLM-5.1-FP8)

Metric	Score
Semantic match	81%
Exact match	48%
GLM ceiling	95%

Apple Silicon

Use the MLX-quantized variant for local inference: rdsm/Qwen3.5-0.8B-parakeet-52k-FullFT-MLX-8bit