Qwen3.5-Omni - Native Audio, Vision and Voice Model

Qwen3.5-Omni is Alibaba's natively multimodal model. It takes text, images, audio, and video as input and produces text plus streaming speech in a single forward pass. Shipped on March 30, 2026 in three variants - Plus, Flash, and Light - it sits at the crossroads of frontier voice AI and open-weight development.

Qwen3-Omni handled 19 input languages for speech and 10 outputs. The 3.5 generation jumps to 113 and 36, which changes the calculus for voice agents outside the English-Chinese axis. The architecture evolves the Thinker-Talker design onto the Hybrid-Attention MoE backbone shared with the rest of the Qwen 3.5 series.

Key Specifications

The 256K context maps to roughly 10 hours of audio or 400 seconds of 720p video at 1 FPS. For meeting transcription or long-form podcast editing, that's enough to process entire sessions in one call.

Qwen3.5-Omni's Thinker-Talker architecture. The Thinker ingests text, audio, images, and video through a Hybrid-Attention MoE. The Talker produces streaming speech tokens concurrently with generation. Source: apidog.com

Benchmark Performance

Alibaba reports 215 SOTA results for Plus across audio, audio-visual, visual, text, and speech-generation benchmarks. The directly comparable numbers against Gemini 3.1 Pro are where the claim carries the most weight.

Word error rate on LibriSpeech clean drops from 3.36 (Gemini 3.1 Pro) to 1.11 - roughly a two-thirds reduction on a benchmark where gains have been fractional for years. Music understanding on RUL-MuchoMusic swings from 59.6 to 72.4, pointing to substantially more musical audio in training than competitors had.

Text performance is the quiet story. MMLU-Redux 94.2 for the Omni variant sits within a point of the non-multimodal Qwen3.5-Plus at 94.3. Adding native audio and video didn't cost the model on text reasoning, which has historically been a tradeoff in unified architectures.

The Artificial Analysis Intelligence Index places Plus at 39 (rank 8 of 67) and Flash at 26 (rank 13 of 78). Both are non-reasoning models.

Key Capabilities

The headline capability is Audio-Visual Vibe Coding: point a camera at a UI or share a screen, describe what you want built verbally, and the model produces functional code from the combined audio-visual stream with no text prompt. The Decoder frames it as emergent rather than an explicit training target - more interesting than the marketing positioning suggests.

Semantic interruption matters for production voice agents. Earlier real-time models struggled to separate a user wanting to cut in from background noise. Qwen3.5-Omni's turn-taking intent recognition distinguishes backchannels ("uh-huh", "right") from actual interruptions. Whether it holds up in a noisy open-plan office is something builders will need to test, but the architecture handles it at the model level rather than via an external VAD pipeline.

Voice cloning works from 10-30 second samples and exposes controls for speed, volume, and emotion. On Seed-zh stability (lower better), Plus scores 1.07 against ElevenLabs 13.08, GPT-Audio 1.11, and Minimax 1.19. On Seed-hard, cloned-voice WER is 6.24 versus ElevenLabs 27.70. That gap matters for customer-facing products where voice drift or word errors kill the experience.

Pricing and Availability

DashScope pricing differs between international (Singapore) and mainland China regions. Headline rates via Artificial Analysis:

At $0.10/M input, Flash is one of the cheapest frontier-tier multimodal APIs - matching text-only Qwen3.5-Flash pricing but with audio, video, and speech output folded in. Plus is still under a third of Gemini 3.1 Pro's rate.

Access paths:

1. DashScope API - Plus and Flash via qwen3.5-omni-plus and qwen3.5-omni-flash model IDs, OpenAI-compatible interface
2. Qwen-Omni-Realtime WebSocket - live audio and video streams with full-duplex conversation and semantic interruption (docs)
3. Hugging Face - Light variant weights plus demo spaces (online, offline)
4. Qwen Chat - consumer-facing test interface at chat.qwen.ai

New DashScope accounts in Singapore get a free quota of 1M input and 1M output tokens for 90 days. Audio tokenizes at roughly 427 tokens per minute, so a 10-minute voice call burns around 4,300 input tokens before any response.

Strengths

• SOTA on directly comparable audio benchmarks - LibriSpeech WER and MMAU ahead of Gemini 3.1 Pro
• Native multimodal single forward pass with streaming speech output concurrent with generation
• 256K context for 10+ hour audio or 400-second 720p video in one call
• 113-language speech recognition and 36-language generation
• Plus pricing undercuts Gemini 3.1 Pro 5x on input and 2.5x on output
• Voice cloning stability far ahead of ElevenLabs on Seed-zh and Seed-hard
• Light variant ships with open weights, preserving Qwen's Apache 2.0 tradition

Weaknesses

• The 215 SOTA claim is thin - many entries are per-language ASR and S2TT subtasks with little competition
• Plus needs 40GB+ VRAM for FP16 self-hosting, pricing individual developers out
• Parameter counts for Flash and Light aren't formally disclosed
• DashScope tiered pricing can get complex on long prompts - chunking may be cheaper
• Plus and Flash are API-only; only Light has published weights
• Non-reasoning model, trails reasoning frontier on chain-of-thought benchmarks
• Verbose outputs - Artificial Analysis measured 16M tokens against a 7.2M average

Related Coverage

• Qwen3.5-Omni Does 10-Hour Audio and 4M Video Frames - original launch coverage
• Alibaba Drops Qwen3.5: 397B Parameters, 17B Active - flagship text model in the same generation
• Qwen3.5-35B-A3B - open-weight sibling sharing the MoE backbone
• Qwen3.5-Flash - text-only hosted API variant at similar pricing
• Gemini 3.1 Pro - closest audio-benchmark peer
• AI Voice and Speech Leaderboard
• Audio Understanding Benchmarks Leaderboard
• How to Set Up an AI Voice Agent
• Qwen 3 Review

Sources

• Qwen3.5-Omni: Scaling Up, Toward Native Omni-Modal AGI - Qwen Blog
• Qwen3.5-Omni Technical Details - Analytics Vidhya
• Qwen3.5-Omni: SOTA in 215 Benchmarks - StableLearn
• What Is Qwen3.5-Omni - WaveSpeedAI
• Qwen3.5-Omni API Pricing - WaveSpeedAI
• Qwen3.5 Omni Plus Price Analysis - Artificial Analysis
• Qwen3.5 Omni Flash Price Analysis - Artificial Analysis
• Qwen3.5-Omni Beats Gemini on Audio - Apidog
• Qwen3.5-Omni Can Hear, Watch, and Clone Your Voice - Decrypt
• Qwen3.5-Omni Review - BuildFastWithAI
• Qwen3.5-Omni Learned to Code from Spoken Instructions - The Decoder
• Qwen3-Omni GitHub Repository
• Qwen-Omni Realtime API Docs - Alibaba Cloud
• Qwen Pricing - PricePerToken