Maxkuo/TEA-ASR-1.1-mlx-q8

TEA-ASR-1.1 MLX 8-bit

8-bit (group size 64) MLX quantization of JacobLinCool/TEA-ASR-1.1 (revision bda08df76d4fd6b487b4a1dd7f0bddf8541696f8) for mlx-qwen3-asr 0.4.4 on Apple Silicon. TEA-ASR-1.1 is a Taiwan Mandarin adaptation of Qwen3-ASR-1.7B. This copy is pinned by revision in a personal Mac app; nothing about the model was changed besides the quantization and the tokenizer file format below.

Usage

import re

import mlx_qwen3_asr as m

result = m.transcribe("audio.wav", model="Maxkuo/TEA-ASR-1.1-mlx-q8", language="Chinese")
text = re.sub("[-]+", "", result.text)

Strip private-use characters (U+E000 to U+F8FF) from the decoded text. TEA-ASR rewrites 12,379 simplified-character tokens as Traditional Chinese and marks 1,161 of them with a trailing private-use character; the upstream tokenizer.json removes those marks in its decoder (a Replace step), but mlx-qwen3-asr decodes from vocab.json and does not.

What was converted

• vocab.json and merges.txt are generated from TEA-ASR's own tokenizer.json (model.vocab, and model.merges joined with a space under a #version: 0.2 header). They differ from stock Qwen3-ASR, so do not swap in the Qwen tokenizer files.
• Weights: loaded with mlx_qwen3_asr.load_models.load_model(dir, dtype=mx.float16), quantized with mlx_qwen3_asr.convert.quantize_model(model, bits=8, group_size=64, encoder_bits=8) (every Linear and Embedding in both the audio encoder and the text decoder at 8 bits), and saved with mx.save_safetensors. lm_head is tied to the token embedding and is not stored; the loader re-ties it. This is the same recipe as moona3k/mlx-qwen3-asr-1.7b-8bit.
• config.json and tokenizer_config.json are copied unchanged from the source revision. quantization_config.json is {"bits": 8, "group_size": 64, "audio_tower_bits": 8}.

Tools: mlx 0.32.4, mlx-qwen3-asr 0.4.4, Python 3.12, macOS on Apple M5.

Files

File	Bytes	SHA-256
weights.safetensors	2169903389	1aea16cc35a10e5d73d58c501a528b350488ff28de8fdd669f9303c523073cb2
config.json	5820	310e42c3f82d422e8e26e582ad53a595b747fa066bb97e33f1dc5478d63849b1
tokenizer_config.json	12517	7f335cbd3b6f1036808e23613c1c1919f8d5570c11cb55420e5094e036f45669
vocab.json	3086754	510171fc05d6cb92892dce8bf9c030263df91252e052e1782d5f3d00464d8eab
merges.txt	1485405	c90018635ffe9024f5424fa9d400deef6eda5a527aac8e5e7ad023055da0faa9
quantization_config.json	60	b365a9bb14e0d066647297dee94dea21df4d7e8fec52356acd23cadfd9ce43e3

License

The TEA-ASR adaptation and checkpoint are released under the MIT License (upstream model card). The underlying Qwen3-ASR weights by Alibaba Cloud remain subject to the Apache License 2.0 and its attribution terms. Accuracy and limitations are described on the upstream model card; please cite TEA-ASR as it asks.