Qwen3.8 Omni Flash: Fed It Images, Video, and Audio All at Once

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Qwen3.8 Omni Flash: Fed It Images, Video, and Audio All at Once · Channel: Fahd Mirza · 09:51 · Watch on YouTube

Topics: AI evaluation and benchmarks, Multimodal models

Summary

The video tests Qwen 3.8 Omni Flash across image understanding, video, audio, multilingual translation, and combined multimodal input. The presenter highlights its detailed visual reasoning, ability to distinguish sampled video frames from actual audio, and accurate analysis of a chemistry diagram alongside an unrelated audio clip. He also notes that the model is comparatively inexpensive for audio and video input, while some multilingual translations appear less accurate than earlier versions. The presenter hopes the model will be open-sourced and plans to cover more of its capabilities in future videos.

Key points

Qwen 3.8 Omni Flash is presented as a native multimodal model for text, images, audio, and video, with a 1 million-token context.
The image test asks whether a woman is wearing a seat belt and what her tattoo symbolizes; the model examines several areas and labels its interpretation of the tattoo’s symbolism as speculative.
The presenter says official benchmarks show strong audio-visual understanding and highlights the model’s lower audio and video input costs compared with Gemini 3.8 Flash and its predecessor.
In a video test, the model correctly recognizes that it has sampled frames rather than actual audio, while describing visual and cultural details.
The presenter describes an architecture combining a vision encoder and an audio-text module with a main agent that can plan and use tools.
An Arabic audio test produces a language identification and translations; the presenter says audio understanding is strong but some translations contain mistakes.
In a combined test, the model accurately describes a chemistry titration curve and transcribes a separate, low-quality English audio clip.
The model states that the chemistry image and audio clip are unrelated instead of inventing a connection.
The presenter says Qwen plans to open-source the model and that further videos will cover its features.

Chapters

Start	Chapter	Summary
00:01	Introducing the model and testing image understanding	The presenter introduces Qwen 3.8 Omni Flash, discusses its benchmarks and pricing, and tests its reasoning about a seat belt and tattoo.
03:08	Video understanding and model architecture	A video test demonstrates the model’s visual descriptions, followed by an overview of its multimodal architecture and possible capabilities.
05:43	Arabic audio and multilingual translation	The presenter tests Arabic audio recognition and translation, praising the audio understanding while noting errors in some translations.
07:47	Combining image and low-quality audio	The final test pairs a chemistry image with a separate audio clip, and the model describes and transcribes them without claiming a false connection.

Quotes

handling text, image, audio, and video in one shot with a 1 million token context. — 00:01

clearly labeled speculative symbolism instead of stating it as a fact. — 02:37

It has correctly read the model as a frame sample video, not real audio. — 04:41

it honestly said that two had no connection instead of forcing a fake link between them. — 09:18

Concepts and entities

Qwen3.8 Omni Flash, Qwen, Hermes Agent, Gemini 3.8 Flash, Fahd Mirza, Happy Horse

Full transcript (original language)

00:01 Quen has struck again, this time with Quen 3.8 Omni Flash, their new native Omniod model, handling text, image, audio, and video in one shot with a 1 million token context. Let’s stop talking about it and start testing it live with Hermes agent where I already have configured it with API. Quen is also going to open source it very very soon. In the first test, what I’m going to do, I’m going to give it this image

00:33 where I’m going to ask the model if this lady is wearing seat belt, right? And not only that, but also what does this tattoo symbolizes? I already have attached the image. And this is my prompt for this image. So, the model is reflecting through. While that happens, let’s talk more about this model. By the way, this is Fahad Miza and I welcome you to the channel. Recently, I also have started this free weekly AI newsletter which you

01:04 can subscribe at fadmiza.substack.com. And these are the official benchmarks from Quen. As you can see, it’s a sweep and the pattern is consistent. Quen 3.8 8 Omni tops or leads most of the audio visual agent and understanding charts, especially the multis speaker meeting transcription test where it blows past every competitor. For me, I think the point here is the price chart at the bottom. Let me see if

01:35 I can show you. There you go. So, look at it. It’s dramatically cheaper on audio and video input than Gemini 3.8 flash and its own predecessor while still winning on quality. The model has come back with the response. But before even I show you this result, I just want to say a few things on its reasoning because I think that is the differentiator now between these huge large scale models. So look at this reasoning. I’m just going to go through it. It didn’t just look once. It

02:06 zoomed into the shoulder, chest, other parts and buckle separately cross checked its own geometry logic and caught itself when I first read some um thing off around before settling on an answer. So not only quite broad but occasionally quite deep reasoning wherever it was needed unlike other previous models from coin where it really wasted lot of tokens but I think this one is quite efficient in that regard.

02:37 Now the actual result. It called the belt technically buckled but flagged real faults strap drifting towards the turnum neck instead of the shoulder and a loose high riding lap belt while giving the tattoo a detailed style breakdown plus clearly labeled speculative symbolism instead of stating it as a fact. A world-class answer. Very very correct. And I think this one has really um passed it with flying color

03:08 especially when it comes to C1. Next up, let’s do a video understanding text with some also audio understanding. Let me play this AI generated video. [music] So there is a music, there’s some dancing, there are some people. Let me take you back to my code. So this is the code which I’m using sending it the prompt the model and also

03:38 um the API key which I’m using from their website. So let me take you back and run this from the terminal. The model is really streaming down the whole response. While it does that, let me quickly walk you through the architecture because I think that is one real good thing happening here. So you can see it’s a vision encoder and special audio text module feeding into a main agent that plans um calls tools and

04:11 orchestrate everything from movie commentary to music video generation. It’s not really just a model that understands audio and video. It’s built to act on them, watch a film and narrate it, translate a drama across languages while preserving each speaker’s voice or take a raw song and turn it into a fully edited music video all from a single prompt. That is why I cannot really do justice to this model in one video. So

04:41 we there will be multiple videos on every feature. So for this video, I just want to give you the flavor of what this model is capable of. Anyway, let’s go back to our terminal. So there you go. Another solid response. I must say it has correctly read the model as a frame sample video, not real audio. So it is quite honest. It reasoned honestly about that limit instead of pretending to hear the music and still really was spot on around

05:11 cultural detail. uh sari colors, jewelry names like mang tikka, gajra, bindi and the marble mogul rajput architecture and sync the clapping to post timing convincingly frame by frame. It even got the happy horse generation watermark on its own. Yeah, that is correct. I generated it with happy horse again from Quen and that video is already on the channel. So solid attention to the detail. Um so only knock I would say that um or maybe it’s a praise really

05:43 that it really it was honest about not hearing audio internally but still writes the final answer as if it heard the music sinking to claps which um we can argue that overstating certaintity in the [clears throat] sorry in the polished output versus its own reasoning but all in all I must say it’s a solid result. And now let’s do multilinguality and audio understanding in the same go. I’m

06:14 going to give it this Arabic audio. Let me play it. And I’m asking the model to tell me which language is that and then translate the sentence to all of these languages. So let me now run this from my terminal and it is giving us the translation. Let’s wait and you have to help me out by checking your language and let me

06:45 know what do you think if it is correct or not. And I have also just translated that Arabic um audio. And the translation in English is subtropical storm Cherry formed today in the Atlantic Ocean as the 10th named storm of the Atlantic hurricane season. If you look at this response, you know what models transcription and translation is correct. Not only it has identified the Arabic and then it has just gone in and

07:16 given us the translation. You see it has even given it here. So I’m just going to scroll through. Please check your uh language and let me know what do you think. I think there are few mistakes here. For example in the naming of the ocean. So I’m just going to scroll through. Let me know check out your language because you know we have short time. I can’t do every language. So some of them

07:47 are correct but there are few mistakes I can tell and that is what I also felt with the other qu 3.8 models that multilinguality has gone down a bit. It was even better in 3.7 anyway. So this is a multilinguality plus audio test. Audio understanding is spoton and in this final test I’m going to combine the modality. So I’m going to give it this science image and we’ll ask it to reason through it what exactly that is. Plus, I’m also going to give it

08:18 an audio in English, a very um how I mean it’s a low quality audio. So, I’m just trying to see if not only it can combine that those modalities, but can it keep it them separate and can it really extract it through a lowquality audio? Let me also play the audio now. However, due to the slow communication channels, styles in the west could lag behind by 25 to 30 years.

08:48 So, it’s a very low quality one. Let’s now run that script. So, there you go. So, you see description of image, transcription, and assessment. That’s correct. I want three things. And if you’re wondering, this is a script which I’m using where all of these two text and image URL and input audio three modalities are being merged together. Another solid answer. It has correctly read that titration curve in real

09:18 chemistry detail, buffer regions, equivalence points and everything accurately um described. Plus the transcription is also accurate which is the unrelated audio clip. And then it honestly said that two had no connection instead of forcing a fake link between them. And that was a hard test. And I’ll just scroll through so that you can also check. You see the image is chemistry and this also look at this. It identified that these are slow communication channels.

09:51 Really good stuff. And this is a transcription. There you go. And there is no meaningful connection between these two. So look, I think Quen has done it again. Really, really want this to get open sourced. As I said, this is not the last video in this series. We’ll be covering it more. But now you have got the flavor of this model. Again, if you want to help out the channel, please become a member. Thank you for all the support.