Qwen3-Omni Technical Report

model 2509.17765
multimodalMoEspeech-generationstreaming-inferenceomni-modelaudio-understanding

Qwen3-Omni Technical Report — L2 #

§1 TL;DR #

Qwen3-Omni: Thinker–Talker MoE architecture unifying text/image/audio/video perception and generation without modality degradation vs unimodal counterparts; open-source SOTA on 32/36 audio benchmarks; 234 ms first-packet speech latency via multi-codebook streaming.


§2 Q1 / Q2 / Q3 #

Q1 — 痛点 #

Contemporary multimodal models exhibit modality trade-offs: gains in one modality degrade performance in others. Before Qwen3-Omni, no open model maintained state-of-the-art across text, image, audio, and video simultaneously. Cascaded pipelines (separate ASR → LLM → TTS) add latency and lose the cross-modal reasoning that Qwen3-Omni's end-to-end design enables.

Q2 — 方法 #

Five architectural innovations over the Thinker–Talker baseline (Qwen2.5-Omni):

  1. MoE for both Thinker (30B-A3B) and Talker (3B-A0.3B) — reduces KV-cache IO and maintains TPS under concurrency.
  2. AuT (Audio Transformer) — replaces Whisper; trained from scratch on 20M hours of supervised audio; 0.6B params; 12.5 Hz token rate with block-wise dynamic window attention (1–8 s).
  3. Multi-codebook RVQ speech representation — hierarchical prediction: backbone → linear head → zeroth codebook → MTP module → residual codebooks.
  4. Causal ConvNet Code2Wav — replaces block-wise DiT vocoder; enables frame-by-frame streaming from first codec token.
  5. Talker decoupling — Talker conditions only on audio/visual features and discrete text tokens, not Thinker's high-level text representations; enables external intervention (RAG, safety filters) on text output.
  6. Training pipeline:

    • Pretraining: three stages — encoder alignment (separate encoder training on frozen LLM) → general multimodal (~2.26T tokens) → long context extension (8K → 32K).
    • Thinker post-training: SFT → strong-to-weak distillation (teacher: Qwen3-32B / Qwen3-235B-A22B, KL minimization) → GSPO (rule-based + model-based rewards).
    • Talker post-training: speech mapping → continual pretraining (hallucination mitigation) → DPO (multilingual preference pairs) → speaker fine-tuning.

    核心技术壁垒: The non-degradation property through early-stage co-training. Mixing unimodal and cross-modal data from the beginning of text pretraining — combined with a deliberate encoder alignment strategy that trains encoders separately on a frozen LLM before joint fine-tuning — achieves multimodal parity. The 20M-hour supervised audio dataset for AuT is an additional data moat practically irreproducible outside Alibaba's scale.

    Q3 — 结果 #

    • Text: parity with Qwen3-30B-A3B unimodal (MMLU-Redux 89.3 vs 89.2; GPQA 70.4 vs 62.9; MultiPL-E 83.8 vs 79.3).
    • Vision: comparable to Qwen2.5-VL-72B; MMMU-Pro 57.0 vs GPT-4o 51.9; MathVista 75.9 vs Qwen2.5-VL-72B 74.8.
    • Audio: open-source SOTA on 32/36 benchmarks; overall SOTA on 22; Librispeech clean WER 1.22 (vs Seed-ASR 1.58, GPT-4o-Transcribe 1.39).
    • Audiovisual: WorldSense 54.0 (vs Gemini-2.5-Flash 50.9); DailyOmni 75.8 / VideoHolmes 57.3 (both SOTA).
    • Speech generation: competitive with dedicated TTS systems; multilingual across 10 languages; cross-lingual competitive with CosyVoice3.
    • Latency: 234 ms cold-start first-packet (audio); RTF 0.47–0.66 at concurrency 1–6.
    • Languages: 119 text, 19 speech-in, 10 speech-out.

    §3 架构 / 方法图 #

    Overall Architecture #

    graph TD subgraph Inputs TEXT["Text — BPE tokenizer, 151K vocab"] AUDIO["Audio — 16 kHz → 128-ch mel-spectrogram"] VISION["Image / Video — dynamic FPS aligned to audio rate"] end subgraph Encoders AUT["AuT Encoder · 0.6B
    Conv2D 8× downsample → attn layers
    dynamic window 1–8 s · output 12.5 Hz"] VIS["Vision Encoder · SigLIP2-So400M, 543M
    image + video trained"] end AUDIO --> AUT VISION --> VIS subgraph Thinker["Thinker — MoE Transformer · 30B total, 3B active"] TMROPE["TM-RoPE — temporal:24 · height:20 · width:20
    absolute temporal anchoring at 80 ms resolution"] TK_LAYERS["MoE Decoder Layers
    chunked prefilling · async with Talker"] TMROPE --> TK_LAYERS end TEXT --> TMROPE AUT --> TMROPE VIS --> TMROPE TK_LAYERS -->|"text tokens (discrete)"| TEXT_OUT["Text Output
    (interceptable by RAG / safety)"] TK_LAYERS -->|"multimodal features + history"| TALKER subgraph TALKER["Talker — MoE Transformer · 3B total, 0.3B active"] TLK_BACKBONE["MoE Backbone"] LINEAR_HEAD["Linear Head → C₀ (zeroth codebook)"] MTP_MODULE["MTP Module · 80M dense transformer
    fixed-step AR · fixed KV cache"] TLK_BACKBONE --> LINEAR_HEAD LINEAR_HEAD --> MTP_MODULE end MTP_MODULE -->|"multi-codebook RVQ tokens · 12.5 Hz"| C2W["Code2Wav · causal ConvNet 200M
    left-context only · frame-by-frame"] C2W --> SPEECH["Speech Waveform — streaming output"]

    Component Parameter Summary #

    ModuleArchitectureTotal ParamsActive ParamsStreaming
    AuT EncoderAttention encoder-decoder650M650M
    Vision EncoderSigLIP2-So400M540M540M
    ThinkerMoE Transformer30B3B
    TalkerMoE Transformer3B0.3B
    MTPDense Transformer80M80M
    Code2WavConvNet200M200M
    Total~34.5B~4.8B

    TM-RoPE Positional Encoding #

    TM-RoPE extends M-RoPE by incorporating absolute temporal information. The 64 rotary angles factorize into 24 temporal + 20 height + 20 width (vs M-RoPE's 16 temporal), redistributing capacity from spatial to temporal to balance local semantics with long-range dependencies. For text, all three components share identical position IDs (equivalent to 1D RoPE). Audio tokens receive temporal IDs at 80 ms resolution. Video frames get monotonically increasing temporal IDs dynamically adjusted to actual timestamps — replacing Qwen2.5-Omni's fixed 2-second chunking with continuous temporal alignment supporting arbitrary-duration inputs.

    Talker Decoupling #

    Qwen3-Omni decouples the Talker from the Thinker's high-level textual representations. The Talker now conditions on: (i) historical discrete text tokens (information-equivalent to embeddings for textual content); (ii) multimodal features from the Thinker (necessary for prosody/timbre coordination in speech translation); (iii) current turn's streamed text. This decoupling enables independent system prompts for response style vs audio style, and allows external modules (RAG, function calling, safety filters) to intervene on the Thinker's textual output before Talker synthesis.

    Streaming Pipeline #

    The streaming design chains four stages with overlapping execution: (1) chunked prefilling — audio/vision encoders output chunks along the temporal dimension; Thinker and Talker perform asynchronous prefilling. (2) Thinker generates text token-by-token. (3) Talker generates one codec frame per step at 12.5 Hz; the MTP module immediately predicts all residual codebooks for that frame. (4) Code2Wav synthesizes the waveform attending only to left context. Unlike Qwen2.5-Omni which requires accumulating a block of Talker output before synthesis, Qwen3-Omni outputs waveform from the very first Talker token.


    §4 作者证明 #

    无形式化作者证明 — 仅实证。The paper contains zero numbered equations; all architecture and methods are described in prose.

    Numerical Consistency Checks #

    #CheckVerificationStatus
    1Parameter countTable 1: 650M + 540M + 30B + 3B + 80M + 200M ≈ 34.5B total; active: 650M + 540M + 3B + 0.3B + 80M + 200M ≈ 4.8B
    2Token rateAuT: Conv2D 8× downsample on 10 ms hop → 80 ms per frame = 12.5 Hz; consistent across §2.2, §2.3, §2.5
    3First-packet latencyTable 2 @1-concurrency: 72 + 88 + 57 + 14 + 3 = 234 ms; matches Abstract claim
    4Pretraining dataS2: 0.57T + 0.77T + 0.82T + 0.05T + 0.05T = 2.26T; consistent with stated "~2 trillion tokens"
    5RTF derivationRTF = (Thinker 1-tok + Talker 1-tok + MTP + codec) / 80 ms; @1-concur: (13.3 + 7.1 + 14 + 3) / 80 ≈ 0.47; matches Table 2
    6Language count119 text + 19 speech-in + 10 speech-out; consistent across Abstract, §1, Table 3, §7

    Parameter Breakdown #

    Thinker: 30B total with 3B active implies ~10:1 ratio (exact expert count, routing mechanism, top-k selection, and auxiliary loss are not disclosed). Talker: 3B total with 0.3B active, same ~10:1 ratio. MTP is a dense (non-MoE) 80M-parameter transformer with fixed KV cache for acceleration.

    Undisclosed architectural detailStatus
    MoE expert count (routed + shared)[论文未披露]
    MoE routing mechanism and top-k[论文未披露]
    Attention head count and hidden dim[论文未披露]
    FFN intermediate dim[论文未披露]
    KV-cache bytes per token[论文未披露]
    Scaling-law fit[论文未披露]
    MoE auxiliary loss formulation[论文未披露]

    §5 实验与数据 #

    Training Recipe #

    StageGoalData (tokens / mix)LR / ScheduleContextKey Techniques
    Pretrain S1: Encoder AlignmentAlign AuT + vision encoder to LLM[论文未披露][论文未披露][论文未披露]Encoders trained separately on frozen LLM; adapter first, then encoder. Joint enc+adapter training abandoned (causes encoder to compensate for frozen LLM limitations)
    Pretrain S2: GeneralFull multimodal capability~2.26T (text 0.57T, audio 0.77T, image 0.82T, video 0.05T, AV 0.05T)[论文未披露]8,192Diverse NL prompts; unimodal + cross-modal data mixed from start
    Pretrain S3: Long ContextExtended sequence understandingLong audio/video heavy mix[论文未披露]32,768Increased proportion of long sequences
    Thinker Post-S1: SFTBridge pretrain → downstreamChatML format, multi-modal[论文未披露][论文未披露]Lightweight SFT; deliberately diverges from pretrain data schema while maintaining architectural consistency
    Thinker Post-S2: DistillationReasoning transferOff-policy → on-policy[论文未披露][论文未披露]Teacher: Qwen3-32B / Qwen3-235B-A22B; student logits aligned via KL divergence minimization
    Thinker Post-S3: GSPOCapability enhancementMulti-modal[论文未披露][论文未披露]Rule-based rewards (math/code/IF with predefined rules) + model-based rewards (Qwen3 judge general + Qwen2.5-VL visual judge)
    Talker S1: MappingMultimodal → speechHundreds of millions of speech samples[论文未披露][论文未披露]Establishes monotonic mapping from multimodal representations to speech
    Talker S2: CPTQuality + long-contextHigh-quality data[论文未披露][论文未披露]Alleviates hallucinations from noisy S1 data; long-context training for extended inputs
    Talker S3: DPOMultilingual stabilityDiverse multilingual speech preference pairs[论文未披露][论文未披露]Preference optimization for multilingual generalization and system stability
    Talker S4: Speaker FTVoice customizationSpeaker-specific data[论文未披露][论文未披露]Refines naturalness, expressiveness, controllability for specific voices

    Hardest-to-replicate training trick: The encoder alignment ordering — training encoders separately on a frozen LLM rather than jointly. The paper explicitly states that joint encoder+adapter training while keeping the LLM frozen causes the encoder to learn compensatory representations for the LLM's limitations, degrading perception. This insight, combined with the 20M-hour AuT dataset, is the primary reproducibility barrier.

    Key Experimental Results #

    Text → Text (Instruct, non-reasoning)

    BenchmarkQwen3-Omni-30B-A3BQwen3-30B-A3B (text-only)Qwen3-235B-A22BGPT-4o-0327
    MMLU-Redux89.389.291.386.8
    GPQA70.462.966.969.7
    AIME2561.324.726.765.9
    ZebraLogic90.037.752.676.1
    WritingBench85.577.075.583.0

    Qwen3-Omni-Instruct surpasses the 8× larger Qwen3-235B-A22B on GPQA, AIME25, ZebraLogic, and WritingBench. Text capability is on par with text-only Qwen3-30B-A3B-Instruct-2507, confirming the non-degradation claim.

    Audio → Text (selected)

    BenchmarkQwen3-OmniSeed-ASRGPT-4o-TranscribeGemini-2.5-Pro
    Librispeech clean (WER)1.221.581.392.89
    Librispeech other (WER)2.482.843.753.56
    Fleurs 19-lang avg (WER)5.334.485.55
    VoiceBench Overall (Thinking)88.889.6 (Pro)
    RUL-MuchoMusic52.049.4
    MMAU77.577.4

    Best-in-class Librispeech results among all models tested. Music understanding (RUL-MuchoMusic 52.0) exceeds Gemini-2.5-Pro and all specialist models.

    Non-Degradation Ablation (Base models, Table 16)

    Controlled comparison with matched parameters, identical text/vision data; sole difference is Omni's additional audio/AV data:

    DomainOmni vs text-onlyOmni vs vision-onlyNotes
    General text (MMLU)81.69 vs 81.24Parity
    Coding (EvalPlus)73.96 vs 69.70Omni wins (+4.3) — surprising
    VQA (MMStar)69.6 vs 67.2Omni wins (+2.4)
    OCR (DocVQA)95.27 vs 95.19Parity
    Video (MVBench)69.50 vs 71.87Omni loses (−2.4) — undiscussed

    Adding audio data consistently improves vision on MMMU (59.33 vs 57.22) and OCR tasks. One anomaly: MVBench drops 71.87 → 69.50, not discussed by the authors.

    Speech Generation (selected)

    BenchmarkQwen3-OmniBest competitor
    Seed-TTS test-en (WER)1.39CosyVoice3: 1.45
    Seed-TTS test-zh (WER)1.07CosyVoice3: 0.71
    Multilingual 10-lang avg WER~1.93MiniMax: ~2.34
    Cross-lingual any-to-ko avg WER5.44CosyVoice3: 9.40

    Competitive with dedicated TTS systems while being part of a unified multimodal model.

    Streaming Latency (Table 2)

    Metric1 Concurrency4 Concurrency6 Concurrency
    First-Packet Latency (Audio/Video)234 / 547 ms728 / 1517 ms1172 / 2284 ms
    Thinker TPS756353
    Talker TPS140125110
    Generation RTF0.470.560.66

    RTF stays below 1.0 at all concurrency levels tested. Latency scales roughly linearly with concurrency — the paper's claim that MoE "ensures prefill latency and TTPT remain largely unaffected" understates the 5× increase at 6-concurrency.

    Thinking Model Caveat #

    The Thinking variant degrades ASR and music understanding: Librispeech clean WER worsens from 1.22 → 2.22, and music micro-F1 scores drop across all benchmarks. Chain-of-thought reasoning introduces hallucinations in perception-dominated tasks — the opposite of the typical "thinking always helps" narrative.


    §6 论证链 #

    StepClaimEvidenceSource
    1Modality trade-off is the central obstacle for multimodal modelsPrior models (GPT-4o, Gemini, Qwen2.5-Omni) show performance degradation when adding modalities§1
    2Thinker–Talker MoE + AuT + multi-codebook addresses the architecture gapMoE reduces active params and KV IO; AuT provides dedicated audio encoding at 12.5 Hz; multi-codebook RVQ captures full acoustic detail; ConvNet Code2Wav enables frame-level streaming§2.1–§2.5, Table 1
    3Early multimodal co-training is the key ingredient for non-degradationControlled ablation: Omni base matches or exceeds text-only and vision-only baselines when audio/AV data is mixed from pretraining start; sole variable is the addition of audio data§6, Table 16
    4SOTA audio performance validates the AuT + 20M-hour data investmentOpen-source SOTA on 32/36 audio benchmarks; overall SOTA on 22; beats Seed-ASR, GPT-4o-Transcribe, Gemini-2.5-Pro on key ASR/music benchmarks§5.1.2, Tables 6–8
    5Streaming multi-codebook design delivers real-time interaction234 ms first-packet latency; RTF consistently below 1.0; single-frame synthesis from first Talker token eliminates block-wait latency§2.5, Table 2
    6Cross-modal reasoning emerges from end-to-end architectureWorldSense 54.0, DailyOmni 75.8, VideoHolmes 57.3 — all SOTA — require joint audio-visual reasoning that cascaded pipelines cannot achieve§5.1.4, Tables 11–12

    Argument weaknesses #

    • Step 3 has a leak: MVBench drops (71.87 → 69.50 in Omni vs vision-only), which the authors do not discuss. This undermines the "zero degradation" narrative for video understanding.
    • Step 4 selectively highlights Instruct results; Thinking variant ASR results (worse than Instruct) are deferred to the appendix without integration into the main argument.
    • Step 5 overstates concurrency robustness: 234 ms at 1-concurrency grows to 1172 ms at 6-concurrency (5×), while the paper characterizes this as "largely unaffected."
    • Overall: the paper provides no formal mathematical framework; all claims rest entirely on empirical benchmarks. The non-degradation claim is an existence proof (one model, one training run) rather than a theoretical guarantee.

    §7 实现 cross-reference #

    Released Artifacts #

    Model weights released under Apache 2.0:

    • Qwen/Qwen3-Omni-30B-A3B (base, Instruct, Thinking variants)
    • Qwen/Qwen3-Omni-30B-A3B-Captioner
    • Flash variants also released

    [实现未公开] — training code, AuT training pipeline, encoder alignment implementation, and data processing pipelines are not open-sourced as of the paper's publication.

    Inference Framework #

    Latency benchmarks in §2.5 / Table 2 are conducted on vLLM (Kwon et al., 2023) with torch.compile and CUDA Graph acceleration. The streaming pipeline assumes chunked-prefilling support. No vLLM model file or configuration is published in the paper.

    关键実装細節 #

    1. AuT dynamic window attention: query attention windows range from 1 to 8 seconds, dynamically sized per layer/head. Short windows support real-time prefill caching for streaming; long windows support offline tasks requiring full-context audio understanding. A fixed window size would force a latency-vs-quality trade-off.
      1. Encoder alignment ordering: the paper explicitly abandoned Qwen2.5-VL / Qwen2.5-Omni's approach of training encoder + adapter jointly while freezing the LLM. Instead, Qwen3-Omni first trains the adapter alone on frozen encoder + frozen LLM, then unfreezes the encoder for joint training. Joint training causes the encoder to learn compensatory representations for the frozen LLM's limitations, degrading general perception quality after unfreezing. This ordering detail is critical for reproducing the non-degradation result.
        1. Talker text decoupling: the Talker receives only discrete text tokens (not the Thinker's high-level continuous representations), plus multimodal features and conversation history. This was an intentional design change from Qwen2.5-Omni, enabling: (a) independent system prompts for response style vs voice style; (b) RAG / function-call / safety-filter interception on Thinker output before Talker synthesis; (c) the theoretical possibility of swapping the Thinker without retraining the Talker.
        2. 核心技术壁垒 (dedicated analysis) #

          The single hardest-to-replicate element is the combination of (i) the 20M-hour supervised audio training dataset for AuT and (ii) the early-stage multimodal co-training recipe. The paper demonstrates that the order of training stages matters (separate encoder alignment before joint training), and that the timing of multimodal data introduction matters (from the beginning of pretraining, not added later). These insights are empirical discoveries requiring large-scale ablation studies that are infeasible for most organizations. The AuT dataset composition (80% Zh/En pseudo-labeled ASR + 10% multilingual ASR + 10% audio understanding) suggests substantial internal data labeling infrastructure.