Original Model Reference
Source: Index-Translate-Omni model card. The tables below reproduce the upstream results for the original models. They are not new benchmark results for this quantized export; see the format-specific verification below for the checks actually performed here.
Text Translation
Table with columns: Model, FLORES COMET22 ↑, WMT24++ COMET22 ↑, WMT26 Judge ↑, MEME ↑| Model | FLORES COMET22 ↑ | WMT24++ COMET22 ↑ | WMT26 Judge ↑ | MEME ↑ |
|---|
| Index-Translate 2B | 0.8655 | 0.8489 | 60.26 | 0.6443 |
| Index-Translate-Omni 2B | 0.8655 | 0.8489 | 60.26 | 0.6443 |
| Index-Translate 9B | 0.8789 | 0.8601 | 75.35 | 0.7387 |
| Index-Translate-Omni 9B | 0.8789 | 0.8601 | 75.35 | 0.7387 |
| Index-Translate 35BA3B (preview) | 0.8794 | 0.8586 | 76.76 | 0.7405 |
| Index-Translate-Omni 35BA3B | 0.8795 | 0.8593 | 78.01 | 0.7405 |
Instruction Following
| Model | instTrans Quality / IFscore ↑ | instTrans_minor Quality / IFscore ↑ / Off-target ↓ | IFMTBench XCOMET-XXL / IFscore ↑ |
|---|---:|---:|---:|---:|
| Index-Translate 2B | 0.5391 / 0.7569 | 0.3050 / 0.6586 / 3.97% | 0.7712 / 0.7584 |
| Index-Translate-Omni 2B | 0.5391 / 0.7569 | 0.3050 / 0.6586 / 3.97% | 0.7712 / 0.7584 |
| Index-Translate 9B | 0.6771 / 0.8209 | 0.5222 / 0.7725 / 3.47% | 0.7957 / 0.8760 |
| Index-Translate-Omni 9B | 0.6771 / 0.8209 | 0.5222 / 0.7725 / 3.47% | 0.7957 / 0.8760 |
| Index-Translate 35BA3B (preview) | 0.6901 / 0.8336 | 0.5151 / 0.7715 / 4.05% | 0.7926 / 0.8991 |
| Index-Translate-Omni 35BA3B | 0.6867 / 0.8371 | 0.5163 / 0.7843 / 3.62% | 0.7934 / 0.9070 |
Original Index-Translate figures come from the published benchmarks. The original 35BA3B reference is the published preview checkpoint.
Speech
CER is a percentage (lower is better). BLEU and chrF2 are higher-is-better. Translation cells show BLEU / chrF2 / target CER.
Table with columns: Model, Chinese ASR CER ↓, English ASR CER ↓, English → Chinese, Chinese → English| Model | Chinese ASR CER ↓ | English ASR CER ↓ | English → Chinese | Chinese → English |
|---|
| Index-Translate-Omni 2B | 13.8113 | 34.2560 | 36.982 / 30.357 / 43.3452 | 30.747 / 51.152 / 48.0670 |
| Index-Translate-Omni 9B | 17.8917 | 14.4097 | 43.507 / 36.312 / 41.3767 | 34.656 / 55.376 / 46.3720 |
| Index-Translate-Omni 35BA3B | 13.0053 | 15.4950 | 40.074 / 32.862 / 44.8097 | 32.411 / 53.452 / 46.1700 |
35BA3B uses one shared audio encoder and connector for all four tasks, with repetition penalty 1.1 for English ASR and the original decoding for the other tasks. Results use the same 30 fixed samples per task. One English → Chinese case reached the 256-token limit with no target translation and remains in the score. Sample counts differ across model sizes.
Image Translation
COMET22 × 100 (higher is better). Each Omni model is compared with its Qwen3.5 baseline on paired evaluated subsets.
Table with columns: Model, MMTIT ↑, FLORES-rendered ↑, Vistra ↑| Model | MMTIT ↑ | FLORES-rendered ↑ | Vistra ↑ |
|---|
| Qwen3.5-2B | 52.164 | 50.187 | 55.326 |
| Index-Translate-Omni 2B | 65.809 | 53.612 | 65.596 |
| Qwen3.5-9B | 73.814 | 67.804 | 68.963 |
| Index-Translate-Omni 9B | 77.604 | 68.438 | 71.815 |
| Qwen3.5-35B-A3B | 75.602 | 70.005 | 71.067 |
| Index-Translate-Omni 35BA3B | 77.011 | 69.891 | 71.482 |
Image scores are retained from the evaluated image configurations and cover the evaluated subsets.
FP8 Export
Quantized omni translation model with language and vision weights, original BF16 audio modules, and a portable runtime.
- Language: FP8 weights with per-channel scales and dynamic per-token FP8 activations (
FP8_DYNAMIC).
- Vision: original FP32 tensors, byte-identical to the assembled source.
- Audio: original BF16 audio encoder and connector, with a portable runtime.
lm_head and token embeddings are not quantized. Linear-attention projections remain in the quantization recipe.
The included JointModel explicitly requests dequantize=True for compressed-tensors. This compatibility path uses BF16 compute after loading; it is not a claim of FP8 kernel acceleration.
The public 35B shared-native audio encoder and decoder profile are preserved. English ASR uses repetition penalty 1.1; Chinese ASR uses the public request-local generation guard with max_new_tokens=256. Audio generation is limited to 256 new tokens.
Loading
Download the entire repository, including audio/ and audio-tokenizer/:
pip install -r requirements-joint.txtpython - <<'PY'from joint_runtime import JointModel model = JointModel('.', device_map='auto')print(model.text_generate([{'role': 'user', 'content': 'Translate into English: 人工智能正在改变内容创作。'}]))print(model.audio_generate('<audio>\nTranscribe the speech and translate it into English.', ['/path/to/audio.wav']))PY
JointModel also supports image_generate(messages) with Transformers image-text messages. The runtime extracts only the required helpers from the supplied bound audio code; it does not require the original training repository. Keep the visual weights in FP32 and the audio weights in BF16.
Verification
Saved-checkpoint reload and text generation passed. The portable runtime also passed text and image-description calls, plus real WAV speech-to-English, English ASR and Chinese ASR calls on an H100. Text perplexity is 38.0772 over 250 held-out tokens using the reference loading mode above. These are small smoke tests, not a comprehensive translation, vision or speech-quality benchmark.
The tensor audit checked 31070 quantized linear weights, 333 byte-identical visual tensors, 529 byte-identical audio tensors, and 263 byte-identical preserved language tensors. Exact outputs and limitations are in validation.json; file hashes are in SHA256SUMS.txt.
This repository does not include GGUF files; GGUF is a separate language/vision export without the end-to-end audio pipeline.
Mirrors