CoolFace
Modelpublic

vanch007/Confucius4-T3PO-mlx-4bit

sourceHugging Faceapache-2.0updated 12h agoView on Hugging Face
0likes
Model Card

Confucius4-T3PO · MLX 4-bit

MLX 4-bit conversion of NetEase Youdao's Confucius4-T3PO, a text-to-text simultaneous translation model for Chinese ↔ English. This repository is an independently converted copy of the original weights, not a new model trained by the uploader. Please refer to the original model card and inference implementation for intended use, limitations and the complete streaming protocol.

Conversion

  • —Source revision: 446e5dcca080740f2c2dc9d06a91ed66a9920410
  • —Converter: mlx-lm 0.31.1 on Apple Silicon, 2026-09-24
  • —Quantization: MLX affine, 4-bit, group size 64 (config.json contains the exact settings)
  • —Command: mlx_lm.convert --hf-path netease-youdao/Confucius4-T3PO --mlx-path ./Confucius4-T3PO-mlx-4bit -q --q-bits 4 --q-group-size 64

The quantized weights are smaller than the original BF16 checkpoint. Translation quality and latency may differ from the original model; no benchmark equivalence is claimed.

Usage

Load with mlx_lm.load("vanch007/Confucius4-T3PO-mlx-4bit") or serve with:

bash
mlx_lm.server --model vanch007/Confucius4-T3PO-mlx-4bit --port 8010

T3PO expects the original chat template and streaming prompt. Maintain STREAMING_HISTORY as committed source¦target§ pairs and CURRENT_INPUT as the uncommitted source buffer. An empty/EOS-only response means WAIT; retain the buffer and retry when more text arrives. A nonempty response means TRANS; emit the segment, append its source–target pair to history exactly once, and clear the committed buffer. Force the final buffer at end of input using a minimum of one generated token or an equivalent first-token EOS suppression. The official streaming example also supports latency modes and terminology constraints.

T3PO accepts text, not audio. Speech translation requires a separate streaming ASR model such as Confucius4-R2T2.

Local validation

The converted checkpoint was loaded with mlx-lm and generated nonempty translations in both directions using the official prompt format:

  • —大家好，今天我们讨论项目进度和下一步计划。 → Hello everyone, today we are discussing the project progress and the next steps.
  • —Good morning. Today we will discuss the project schedule and next steps. → 早上好。今天我们将讨论项目进度和后续步骤。

Incremental Chinese input also produced WAIT events before a committed translation. These are functional smoke tests, not a translation-quality benchmark.

License and attribution

The source model is released under Apache-2.0 by NetEase Youdao. The Qwen base model and tokenizer have their own applicable license terms; see the source model card. Credit for the model and streaming protocol belongs to the original authors.