Qwen3.5-Omni is here! Scaling up to a Native Omni-modal AGI

Multimodal AI has grown from novelty to a should in latest occasions. Want proof? If I had been to let you know to work on an AI mannequin that solely understands textual content, you’ll in all probability snicker and throw 10 mannequin names at me that may work throughout codecs – be it textual content, audio, or visuals. The brand new race, thus, for the bigwigs, is to not make simply one other AI mannequin, however a system that may perceive the world extra like people do. That is naturally finished by way of language, visuals, sound, and movement collectively. That’s the house Alibaba’s new Qwen3.5-Omni enters.

The most recent mannequin in Alibaba’s Qwen household is positioned as a “absolutely omni-modal LLM”. We will discover what meaning in idea, and what this moniker guarantees within the practicality of issues. One factor is for positive (with Qwen and different launches just like the latest Gemini 3.1 Flash Dwell), AI fashions have gotten much less like separate instruments and extra like unified interactive methods.

For now, we deal with the Qwen3.5-Omni and all that it brings ot the desk.

What’s Qwen3.5-Omni?

As I discussed earlier, this one is a totally omni-modal mannequin below the Qwen household. In easy phrases, it’s constructed to deal with textual content, pictures, audio, and audio-visual content material inside a single system. That’s what separates it from older AI setups, the place every modality usually wanted a distinct mannequin or pipeline.

As is clear from its launch transient, Alibaba is pitching Qwen3.5-Omni as a mannequin designed for richer, extra pure interplay with real-world inputs. As an alternative of treating voice, pictures, and video as non-obligatory add-ons, it presents them as core elements of the mannequin itself. Which means Qwen3.5-Omni is far more than a typical chatbot. It’s a multimodal AI system meant to interpret totally different sorts of data collectively.

As for its variants, the brand new Qwen3.5-Omni sequence consists of Instruct variants in three sizes – Plus, Flash, and Gentle. This household construction makes it best for various use circumstances and efficiency wants. The launch additionally highlights long-context help, which suggests the mannequin just isn’t solely broad in modality but in addition constructed for heavier, extra sustained inputs.

There are, after all, extra such options in line. Right here is all that the brand new Qwen3.5-Omni brings to the desk.

Qwen3.5-Omni Options

Qwen3.5-Omni is clearly a extra succesful step up from Qwen3-Omni. Although the factor to notice right here is that it comes with a lot broader horizons as properly. Right here is how:

1. Stronger multilingual capabilities

In contrast with Qwen3-Omni, Qwen3.5-Omni comes with considerably improved multilingual capabilities, together with speech recognition in 113 languages.

2. Lengthy-context help

The Qwen3.5-Omni sequence consists of Instruct variations with help for 256K long-context enter. This factors to a mannequin that’s designed for a lot bigger and extra sustained prompts than a typical chatbot workflow.

3. A number of mannequin sizes

The sequence consists of three Instruct sizes: Plus, Flash, and Gentle. That provides Qwen3.5-Omni a extra versatile product household moderately than a single one-size-fits-all launch.

4. Giant multimodal enter capability

The announcement blog says the mannequin can deal with greater than 10 hours of audio enter and over 400 seconds of 720p audio-visual enter at 1 FPS. That signifies that it’s constructed for heavier audio and video understanding workloads.

5. Semantic interruption help

Qwen3.5-Omni helps semantic interruption by way of “native turn-taking intent recognition.” In easy phrases, this helps the mannequin distinguish between significant consumer interruption and irrelevant background noise. All in all, the characteristic makes stay conversations really feel extra pure.

6. Native WebSearch and Operate Calling

The mannequin natively helps WebSearch and sophisticated FunctionCall capabilities. This enables it to determine by itself whether or not it ought to invoke WebSearch with the intention to reply a consumer’s real-time query. Assume extra agent-like sensible use.

7. Finish-to-end voice management and dialogue

This can be a very fascinating improve with Qwen3.5-Omni, and I’m positive you’ll find it irresistible too when you see its demos. The brand new Qwen mannequin helps end-to-end voice management and dialogue. This implies the mannequin can comply with spoken directions in a extra human-like method by controlling facets of speech reminiscent of quantity, pace, and emotion. As demonstrated in some movies, the mannequin can whisper, shout, and even categorical feelings in a method that may sound very pure to most.

8. Voice cloning

One other notable characteristic is voice cloning, which permits customers to add a voice and customise the AI assistant’s output voice accordingly. It means now you can converse to the AI and have it reply within the voice of your selection.

With all these options, right here is how the Qwen3.5-Omni performs in benchmark assessments.

Qwen3.5-Omni: Benchmark Efficiency

Fairly than profitable each single benchmark outright, Qwen3.5-Omni-Plus comes throughout as a really well-rounded omni-modal mannequin that stays extremely aggressive throughout textual content, imaginative and prescient, audio, audio-visual understanding, and speech era. That’s the larger takeaway right here: consistency throughout virtually each format. And as an add-on, it both leads or comes extraordinarily shut usually occasions to the highest mannequin within the comparability.

1. Audio: USP of the mannequin

Audio is clearly one in every of Qwen3.5-Omni-Plus’s strongest areas.

In audio understanding, it barely edges out Gemini-3.1-Professional on MMAU (82.2 vs 81.1) and MMSU (82.8 vs 81.3), whereas additionally delivering a giant leap on RUL-MuchoMusic (72.4 vs 59.6). On dialogue-heavy duties, it posts one of the best rating on VoiceBench (93.1), forward of Gemini-3.1-Professional’s 88.9.

Its speech-related efficiency can be spectacular in transcription and recognition-style duties. For instance, on LibriSpeech, Qwen3.5-Omni-Plus scores 1.11 / 2.23, forward of Gemini-3.1-Professional’s 3.36 / 4.41, and on CV15 (en) it information 4.83 in opposition to Gemini’s 8.73. That implies Qwen is especially sturdy not simply at listening to audio, however at processing it precisely.

2. Audio-Visible: Sturdy, however not all the time the outright chief

On audio-visual duties, Qwen3.5-Omni-Plus performs strongly, although that is one space the place Gemini-3.1-Professional nonetheless holds some benefits.

As an illustration, Qwen leads on DailyOmni (84.6 vs 82.7) and QualcommInteractive (68.5 vs 66.2), and likewise tops Omni-Cloze (64.8 vs 57.2) in captioning. However Gemini stays forward on benchmarks like WorldSense (65.5 vs 62.8), VideoMME with audio (89.0 vs 83.7), and OmniGAIA software use (68.9 vs 57.2).

3. Visible: Aggressive, with some category-leading scores

In visible duties, Qwen3.5-Omni-Plus once more appears to be like balanced and succesful moderately than wildly dominant.

It posts one of the best rating on MMMU-Professional (73.9), edges forward on RealWorldQA (84.1), leads on CC-OCR (83.4), tops EmbSpatialBench (85.4), and performs finest on a number of video benchmarks, together with VideoMME with out subtitles (81.9), MLVU (86.8), MVBench (79.0), LVBench (71.2), and MME-VideoOCR (77.0).

That mentioned, the non-thinking Qwen3.5-Plus baseline nonetheless beats it on some basic visible and reasoning-heavy benchmarks reminiscent of MMMU, MathVision, and Mathvista mini. So Qwen3.5-Omni-Plus is probably not the very best visible mannequin in isolation. Although it nonetheless demonstrates very strong visible efficiency whereas bringing audio and speech into the identical system.

4. Textual content: Strong, however not the headline story

Qwen3.5-Omni-Plus exhibits textual content efficiency, although it doesn’t seem like the central headline of the discharge.

Qwen3.5-Omni-Plus stays near the non-thinking Qwen3.5-Plus mannequin on a number of benchmarks: MMLU-Redux (94.2 vs 94.3), C-Eval (92.0 vs 92.3), and IFEval (89.7 vs 89.7). It additionally does fairly properly on long-context duties like LongBench v2 (59.6) and reasoning duties like HMMT Nov 25 (84.4).

The broader sample is that Qwen3.5-Omni-Plus preserves a powerful textual content basis whereas extending into different modalities. After all, it’s not essentially the most thrilling a part of the benchmark desk. However it’s reassuring that the multimodal enlargement does minimize down on text-quality.

5. Speech Era: Standout benchmark outcomes

This is among the clearest strengths of the mannequin.

In customized voice stability, decrease is best, and Qwen3.5-Omni-Plus performs extraordinarily properly. It scores 1.07 on Seed-zh, beating ElevenLabs (13.08), Gemini-2.5 Professional (2.42), GPT-Audio (1.11), and Minimax (1.19). It additionally leads on Seed-hard (6.24) and performs finest on the multilingual averages proven, together with 2.06 on Public-Multilingual-avg (20 languages) and 5.82 on Inhouse-Multilingual-avg (9 languages).

On voice clone stability, it additionally posts one of the best multilingual rating within the public setting at 1.87, forward of ElevenLabs (10.29) and Minimax (2.52). On voice clone similarity, greater is best, and Qwen3.5-Omni-Plus reaches 0.79 and 0.80, which is once more the strongest rating within the comparability proven.

This makes speech era probably the most compelling elements of the Qwen3.5-Omni-Plus benchmark story.

Total takeaway

• Strongest: Audio, Speech era
• Very sturdy: Audio-visual, Imaginative and prescient
• Strong/above common: Textual content

This efficiency is made attainable due to the distinctive structure of the Qwen mannequin. Right here is

Qwen3.5-Omni Structure

Qwen3.5-Omni follows what Qwen calls a Thinker-Talker structure. Now we have seen it earlier than within the earlier Qwen fashions. As an alternative of treating understanding and response era as one blended course of, the mannequin separates them into two useful elements. That makes the structure simpler to know, particularly for a mannequin constructed to deal with a number of modalities.

Here’s what each elements do:

1. The Thinker

The Thinker is chargeable for the mannequin’s understanding layer. In response to Qwen, it receives visible and audio alerts by way of the mannequin’s encoders and handles the higher-level reasoning over these inputs.

In easy phrases, that is the a part of the system that interprets what the mannequin is seeing, listening to, or studying earlier than a response is generated.

2. The Talker

The Talker handles the output aspect of the system. As soon as the mannequin has processed the enter, this part is chargeable for producing the response.

This distinction issues as a result of Qwen3.5-Omni is not only meant to analyse inputs. It’s also meant to answer interactive and conversational use circumstances.

3. Hybrid-Consideration MoE in Each Elements

Qwen says that each the Thinker and the Talker undertake Hybrid-Consideration MoE.

That element suggests the structure is designed to stability functionality and effectivity. As an alternative of counting on one giant block to handle all the pieces, the mannequin makes use of a extra structured design to help each multimodal understanding and response era.

Why This Structure Issues

For an omni-modal mannequin, structure issues greater than traditional. Qwen3.5-Omni is predicted to course of textual content, pictures, audio, and audio-visual content material inside one system. A break up between understanding and era helps help that broader position.

That is additionally why, moderately than trying like a textual content mannequin with just a few added multimodal options, Qwen3.5-Omni is being framed as a system designed from the bottom up for richer interplay throughout totally different enter and output modes.

Now that we all know the way it works, right here is entry the brand new Qwen mannequin.

Qwen3.5-Omni: Entry

There are 3 foremost methods to entry the Qwen3.5-Omni, largely based mostly in your use case. These are:

1. Qwen Chat

Essentially the most simple strategy to strive Qwen3.5-Omni is thru Qwen Chat, which acts because the direct user-facing entry level for the mannequin household.

Greatest for: particular person customers

2. by way of Offline API in Alibaba Cloud Mannequin Studio

For normal API-based integration, Alibaba Cloud gives Qwen-Omni by way of Mannequin Studio. The mannequin accepts textual content mixed with one different modality right here, reminiscent of picture, audio, or video, and might generate responses in textual content or speech. Alibaba notes that Qwen-Omni at present helps OpenAI-compatible calls solely, requires an API key, and works with the newest SDK.

Greatest for: app integration and multimodal era workflows

3. by way of Realtime API for stay audio and video interactions

For interactive functions, Alibaba Cloud additionally provides Qwen-Omni-Realtime, which is accessed by way of a stateful WebSocket connection. This route is supposed for real-time audio and video chat use circumstances, the place the mannequin can course of streaming inputs and generate responses constantly throughout a session.

Greatest for: voice- or video-driven stay experiences

Qwen3.5-Omni: Demonstration

The Qwen group has shared a number of demos of the brand new Qwen3.5-Omni that showcase its capabilities throughout use circumstances. Examine them out beneath:

1. Audio-Visible Captioning

The primary demo for the mannequin is that of audio-visual captioning. The demo exhibits how the mannequin is ready to precisely interpret the data being shared inside a video and generate the textual content for a similar. Test it out in motion within the embed beneath.

  

2. Audio-Visible Vibe Coding

This one is tremendous fascinating, because it exhibits Qwen3.5-Omni decoding particular technical directions shared inside a video, after which appearing accordingly. As could be seen, the mannequin can clearly perceive what is occurring throughout visible and audio inputs and help in producing or refining code accordingly. That is combining multimodal context into the coding loop, making the interplay really feel extra intuitive than a plain text-only workflow.

  

3. Multi-Flip Dialogue and Clever Interruption

Alibaba additionally shares proof for its claims of multi-turn dialogue capabilities on the Qwen3.5-Omni. In one other video, the mannequin could be seen dealing with interruptions tremendous intelligently. It showcases that the Qwen3.5-Omni can casually maintain a back-and-forth dialog whereas additionally recognising when a consumer is meaningfully interrupting, as a substitute of reacting awkwardly to each sound or pause.

The anchor could be clearly seen attempting to idiot the mannequin with filler phrases like “hmmm” and “okay” in the course of the mannequin’s response. Although Qwen3.5-Omni appears to know higher than to interrupt.

  

4. Voice Model, Emotion, and Quantity Management

Should you had been to ask me, this appears to be the USP of the brand new Qwen mannequin. Now we have all seen AI fashions conversate with us in a really comparable (if not precise) tone as people. The Qwen3.5-Omni now takes it a step additional and brings in voice type, emotion, and quantity management. The demo highlights how the mannequin can whisper, shout, and even narrate a poem whereas feeling dejected. That’s one thing you don’t see too usually.

  

Conclusion

From what we will see within the demos and the data shared by Alibaba, the brand new Qwen3.5-Omni takes the multi-modal capabilities of an LLM to a different stage. From deciphering audio-visual directions to creating AI conversations really feel far more human, it brings with it a set of options which might be not often seen in AI fashions.

I’m positive many would love to change to Qwen3.5-Omni after this, largely for your complete conversations occurring in audio-visual inputs and outputs. Whether or not they ship on the standard that’s showcased right here, stays to be seen.