Model overview
Table with columns: Item, Value| Item | Value |
|---|
| Input | Audio, text, image, and video |
| Output | Text |
| Architecture | Multimodal Mixture-of-Experts |
| Thinker LM | 30B total / 3B activated MoE |
| Primary evaluation language | English |
| Weights | BF16 Safetensors |
Post-training primarily targets speech-conditioned interaction; image and video
capabilities are inherited from the base model and were not comprehensively
re-evaluated.
Training
BR-Voice-Reasoner is trained from Qwen3-Omni-30B-A3B-Thinking using
cross-modal on-policy distillation. During training, the student processes
spoken requests, while a frozen teacher uses their aligned text forms to provide
token-level learning signals along student-generated trajectories.
Optimization is restricted to two audio-language projection modules. The
language model, vision encoder, and remaining audio encoder parameters stay
frozen, focusing post-training on the audio-language interface used to access
the model's existing knowledge, reasoning, and instruction-following
capabilities.
Training data
Training data cover spoken knowledge, reasoning, instruction following,
open-ended interaction, and safety-oriented tasks using both real-world and
synthesized speech.
Evaluation
BR-Voice-Reasoner is evaluated on nine VoiceBench subsets covering knowledge,
reasoning, instruction following, safety, and open-ended spoken interaction.
All scores are reported on a 0–100 scale; higher is better.
Qwen3-Omni results are taken from the
official Qwen3-Omni model card,
while Nemotron 3 Nano Omni results are taken from its
official model card. BR-Voice-Reasoner results
were obtained using the evaluation protocol described below. BR-Voice-Reasoner
values are rounded to the same one-decimal format. The external columns and
BR-Voice-Reasoner were not produced by a single shared evaluation run;
cross-column differences are therefore reported as references rather than as a
strictly controlled comparison.
BR-Voice-Reasoner evaluation protocol
Table with columns: Item, Setting| Item | Setting |
|---|
| Reasoning | Enabled |
| Generation | temperature 0.6, top-p 0.95, top-k 20 |
| Model judge | GPT-4o-mini, three judgments |
| Overall | Mean of nine normalized subset scores |
AlpacaEval, CommonEval, and WildVoice ratings are normalized to a 0–100 scale
before aggregation. The VoiceBench average is the arithmetic mean of the nine
normalized subset scores.
Full generation, seeding, evaluator, and retry settings are provided in
evaluation_protocol.json.
Installation
pip install "transformers==5.12.1" "accelerate" "qwen-omni-utils==0.0.9"
ffmpeg must also be available on the system for media loading. FlashAttention
2 is optional; install it separately with
pip install flash-attn --no-build-isolation on compatible hardware.
Usage
from transformers import ( Qwen3OmniMoeForConditionalGeneration, Qwen3OmniMoeProcessor,)from qwen_omni_utils import process_mm_info MODEL_ID = "brgroup/BR-Voice-Reasoner" model = Qwen3OmniMoeForConditionalGeneration.from_pretrained( MODEL_ID, dtype="auto", device_map="auto", # attn_implementation="flash_attention_2", # optional)processor = Qwen3OmniMoeProcessor.from_pretrained(MODEL_ID) conversation = [ { "role": "user", "content": [ {"type": "audio", "audio": "example.wav"}, {"type": "text", "text": "Answer the question in the audio."}, ], }] prompt = processor.apply_chat_template( conversation, add_generation_prompt=True, tokenize=False,)audios, images, videos = process_mm_info(conversation)inputs = processor( text=prompt, audio=audios, images=images, videos=videos, return_tensors="pt", padding=True,)inputs = inputs.to(model.device).to(model.dtype) text_ids, _ = model.generate( **inputs, return_audio=False, thinker_return_dict_in_generate=True, max_new_tokens=2048,)response = processor.batch_decode( text_ids.sequences[:, inputs["input_ids"].shape[1]:], skip_special_tokens=True, clean_up_tokenization_spaces=False,)print(response[0])
Limitations
- Primary training and evaluation are English-focused; multilingual, visual,
and video capabilities inherited from the base model were not comprehensively
re-evaluated.
- Performance may vary across accents, recording conditions, noise levels,
domains, and prompting styles.
- The model may mishear speech, hallucinate, reason incorrectly, or produce
unsafe or biased content.
- Model-based evaluation results may vary with evaluator versions and inference
settings.
License and attribution
BR-Voice-Reasoner is released under the Apache License 2.0 and is derived from
Qwen3-Omni-30B-A3B-Thinking,
whose original model copyright is Copyright 2025 Alibaba Cloud.
Modifications are Copyright 2026 Bairong Inc. See LICENSE and
NOTICE for license terms and attribution.