Qwen3.5-Omni Technical Report

What the report is

The Qwen3.5-Omni Technical Report is credited to the Qwen Team at Alibaba and was submitted to arXiv on 17 April 2026 under the identifier 2604.15804^[1]. It documents Qwen3.5-Omni, the latest model in the Qwen-Omni family — a line of models built to handle more than one kind of media inside a single system rather than as separate text, speech and vision pipelines^[1]. The model described in the report is scaled to hundreds of billions of parameters and carries a 256K context length^[1].

How it is built

Two ingredients dominate the report: data and architecture.

On the data side, the model is trained on a massive dataset of heterogeneous text-vision pairs together with more than 100 million hours of audio-visual content^[1]. The pairing of text with vision, and the sheer volume of audio and video, is what gives the model its grounding in material that is not text: speech, sound, moving images and the way they line up in time.

On the architecture side, both of the model's main components use a Hybrid Attention Mixture-of-Experts framework^[1]. The two components are the Thinker, which handles understanding and reasoning, and the Talker, which handles speech generation^[1]. A mixture-of-experts layer routes each token to a small subset of the network's parameters, so the total parameter count can be very large while the computation spent per token stays comparatively modest; the report presents this design, combined with its attention scheme, as what enables efficient long-sequence inference^[1]. That matters because audio and video inputs are long by nature — a conversation is minutes, a recording can be hours.

What it can do

The input side is deliberately long-form. The model supports more than 10 hours of audio understanding and 400 seconds of 720P video at 1 frame per second^[1]. Understanding and generation are multilingual: the model covers multilingual understanding and speech generation across 10 languages^[1].

Speech output brings its own difficulty. Streaming speech synthesis — producing audio while the user is still listening — has been unstable and unnatural when the tokenizer used for text and the one used for speech encode at mismatched efficiency^[1]. The report introduces ARIA, which dynamically aligns text and speech units to correct exactly that problem^[1].

Beyond answering, the model produces script-level structured captions with temporal synchronisation and automatic scene segmentation^[1]. The authors also report an ability they call Audio-Visual Vibe Coding: writing code directly from audio-visual instructions^[1]. They describe it as emergent, meaning it was not the stated target of training but appeared in the finished model^[1].

Results

The Qwen3.5-Omni-plus configuration reports state-of-the-art results across 215 audio and audio-visual understanding, reasoning and interaction subtasks and benchmarks^[1]. On key audio tasks it surpasses Gemini-3.1 Pro, and on comprehensive audio-visual understanding it matches it^[1].

Why it matters

The report is a useful marker of where multimodal models were heading: away from attaching a speech module to a text model and toward one system that takes hours of audio and minutes of video in, and streams speech out, in 10 languages^[1]. The engineering choices it highlights — mixture-of-experts for long sequences, and alignment between text and speech tokenizers for stable streaming speech — are the constraints anyone building a real-time spoken assistant runs into^[1]. The claim of an emergent coding ability from audio-visual instructions is the more open-ended part: it suggests that training a model broadly enough on paired sound, image and text can produce behaviour that was not explicitly designed in^[1].