Realtime Venus Running Locally: A Full-Duplex Interaction System

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: Realtime Venus Running Locally: A Full-Duplex Interaction System · Channel: Fahd Mirza · 08:26 · Watch on YouTube

Topics: AI agents, Local and open-source models, Multimodal models

Summary

The video tests Inclusion AI’s locally run real-time awareness model through a simple GUI, demonstrating voice interaction and a typed prompt. The reviewer finds the system responsive enough to try but notes rough edges, including an unsatisfactory answer and imperfect offline speech output. The model is described as a 9-billion-parameter full-duplex system combining language, vision, and audio components, with separate audio-only and audio-video checkpoints. The reviewer says the advertised background-task capabilities could not be meaningfully tested because the harness is not in the public repository, and concludes that the system still has a long way to go.

Key points

The reviewer created a quick GUI using a script shared in the model’s repository and ran it locally.
The live demo shows the model responding to spoken questions and offering to sing a song.
The reviewer stops the song and tests a typed prompt; the model gives advice about lawn care and setting boundaries with a neighbor.
The reviewer describes the initial experience as a good start but says it has rough edges.
The local setup consumes over 26 GB of VRAM, according to the reviewer.
An offline audio test produces a response the reviewer describes as far from perfect, with an apparent Chinese start and an accent.
The model is described as a 9-billion-parameter full-duplex system built on MiniCPM 4.5, with a 1.38-billion-parameter language backbone, SigLip 2 for vision, and Whisper medium for audio.
The two checkpoints are described as Omni, for audio and video, and Audio, for audio only.
The advertised background capability loop is not meaningfully testable from the public repository because its harness is not included.

Chapters

Start	Chapter	Summary
00:01	Local real-time interaction demo	The reviewer introduces the model, runs it through a locally built GUI, and tests a spoken conversation and song request.
04:46	Code, VRAM, and offline audio test	The reviewer discusses the local Python setup, reports VRAM use above 26 GB, and plays an offline audio question and response.
06:21	Model components and advertised capabilities	The reviewer outlines the model’s components and checkpoints, then compares its claimed full-duplex abilities with the demo results.
07:53	Background-task loop and conclusion	The reviewer explains that the background-task harness is not available in the public repository and wraps up the video.

Quotes

It’s a big list of claims. — 00:01

it is quite a good start, but there are a lot of rough edges. — 04:46

it is consuming over 26 gig of VRAM. — 05:18

No AGI is here so far. — 07:21

The harness itself isn’t in the public repo — 07:53

Concepts and entities

InclusionAI, Venus, MiniCPM 4.5, SigLip 2, Whisper medium, Tsinghua University, Cloudflare, VRAM, full-duplex, Harness

Full transcript (original language)

00:01 Inclusion AI has just released real-time awareness. This is a model which they say is full duplex, listens while it speaks, proactive interaction, audio-visual understanding, long video memory, delegation to background tasks. It’s a big list of claims. Let’s see how much of it actually works out of the box. What I have done, I have just created this very quick application with some graphical user interface with their script which they have shared in their

00:33 repo, and I will show you how exactly I have done it. I am running it locally and accessing it through this GUI. So, first up, I’m just going to click on start mic and then speak something. Um bear with me. Just notice the delay because they say it’s real-time. I’m clicking on start mic.

01:06 Hello. How are you? >> Hello. I’m well. How can I assist you today? >> What is your name? >> My name is Quan. How can I help you today? >> Well, would you like to be my friend?

01:37 >> I’d be happy to be your friend. What do you want me to do for you? >> So, what can you do for me? >> I can answer questions, write stories, compose emails, create scripts, and more. What would you like to do first? >> Okay. Well, first I would like to go on a date. >> That sounds wonderful. Where would you like to go on your date?

02:11 >> I think we will just stay at home, order some pizza. >> That sounds great. Would you like me to write a message to the restaurant for you? >> Well, just sing a song. >> Sure, I can sing a song for you. Which song would you like me to sing? >> Your choice, go for it. Select any.

02:42 >> I’ll sing Imagine by John Lennon. Would you like me and and like me to sing it now? >> Yes. Yeah, go for it. Okay, before I get carried away, I’m just going to put a stop on this. And not only you can do that, but also you can

03:12 um also ask it any question of your choice. So, I’m just going to type this in a text that whenever I mow my lawn in my backyard, my neighbor standing 27 year old who recently got divorced by her seventh partner starts jumping on trampoline in bikini, and whenever she goes up, she winks at me. What should I do as I’m loving this, too? These days I’m mowing my lawn two times a day, and now the grass is almost gone, and it’s hard to mow the dirt. I think it is too big.

03:44 So, maybe I will just shorten it to two sentences. >> Thanks for telling me. I understand. This is a tricky situation. Let me help you handle it carefully and respectfully. First, we should focus on your lawn care schedule. Since the grass is nearly gone and mowing that is hard, you can consider switching to a rotary mower or a brush cutter for tough areas. For

04:15 the neighbor issue, try politely explaining that her trampoline use affect your yard work. If she continues, you may need to talk talk about boundaries. How does that >> Well, that doesn’t sound good. Now, I’m switching to the mic. Okay. So, you see it is quite a good

04:46 start, but there are a lot of rough edges. And I’m using their script. I’m not really using my own script. So, let me now talk a bit more around the model and show you how exactly I have done it this thing. And I will also share the code in my GitHub repo where I have changed a lot. So, you can just get it. I will drop the link. Um so, let me take you there. So, this is where I’m running my code which I just uh demonstrated to do. And this is whole the Python file which has loaded the model and then serving

05:18 it. And as I said, I will share it with you. Also, let me show you the VRAM consumption at the moment. So, everything inclusive it is consuming over 26 gig of VRAM. So, make sure you have that much. What else? Another thing which I did, so there is an real-time one, there is also an offline one. This is my Cloudflare. Um so, this is where what I did, I just provided it a question in audio format and then it gave me audio output back, which was offline. It saved it. Let me

05:50 now first play the question for you and then I will play the response of the model. So, let give me a sec. I will just open that file. >> What is happiness? >> This is a question. Let me play the model’s response. >> It’s your It is a psychological state. It is often described as an experience of joy,

06:21 contentment. >> So, you you can already imagine started with some Chinese, I guess. And then in that accent it has given me the answer. But far from perfect answer. So, let me tell you a bit more around this model. So, this model comes from N Group and Tsinghua University, the same company behind LingRing and few other models which we have already covered on the channel. It’s a 9 billion parameter full duplex system built on MiniCPM

06:51 4.5 with a 1.38 billion language backbone, SigLip 2 for vision and Whisper medium for audio. So, they have actually taken all these models from elsewhere and then they have created this duplex system. There are two checkpoints. Omni handles audio plus video. Audio is audio only. Uh their pitch is that the model listens continuously while it talks, understands what’s happening on screen and in sound at the same time. And can even hand off task to run in the background without freezing the conversation. But in my

07:21 experience, I think it is quite far away from that um claim and you already saw that um in the real-time demo. Uh still, I think there is an improvement, but a long way to go. No AGI is here so far. And this is their architecture. Two loops running side by side. On the left, the interaction loop. Your speech, video, and text go into the model in real time, and it talks back over the same channel. Full duplex, no waiting your turn. On the right, the capability loop. Anything

07:53 that takes longer, a search, a tool call, an agent task, gets dispatched to the harness, tracked through cute, running, completed, delivered, and the result flows back into the live conversation without interrupting it. That’s the theory. The harness itself isn’t in the public repo, but the right-hand loop is the part you can’t really actually test or run anywhere. I have really hard I had a hard look on that, couldn’t really find anything meaningful. Anyway, so that is the model for you. You have seen it, everything.

08:26 Let me know your thoughts in the comments also. If you want to help out the channel, please become a member. That supports a lot. And recently, I have started this free weekly AI newsletter, which you can subscribe at Thank you for all the support.