AI never sleeps and this week has been absolutely insane. ByteDance drops an open-source video editor just like Gemini Omni. Google releases a really
cool open-source real-time music generator. ChatGPT is now able to dream. Nvidia releases their best model yet and it's fully open-source. They also
released a super realistic world model which is also open-source. We also have a new frontier open
model from Minimax. We have not one but two frontier image models. Google drops a super efficient model that can run offline on just your laptop.
Baidu drops a new video model with audio natively baked in. Alibaba drops a real-time video generator which you can stream and edit with prompts in
real-time. We have some ridiculous humanoid robot demos and a lot more. So, let's jump right in. Thanks to
HubSpot for sponsoring this video. First up, ByteDance releases a really flexible video model called Bernini. This is kind of like an open-source
Gemini Omni. So, here they call it a unified model that can edit videos with a text prompt or image or video references. For example, we can take this
existing video on the left and then add two characters to the video plus change it to a warm golden
tone. Or here, we can remove the left sheep. Or we can change the perspective into an aerial view like this. Or we can also easily replace the
background with something else. Or completely change the scene into winter. I was especially impressed with this demo where it's able to shift the
focus to the background instead of the foreground. Now, what I think is even more powerful is you can input a
reference image and insert that reference into an existing video. For example, we can turn the flowers into the flowers from the image and here's what
we get. Or we can add this helicopter into the video like this. Or we can also completely replace the background with a reference image. And the
reference image can also act as the artistic style of the video. So, for example, we can turn this
original real video into this flat animated style. We can also change the weather of the video to match an input image. So, this understands text,
image, and videos as references, making it very flexible. So, here's a cool example where we can insert the video on the left into the screen of the
second video. And here's our result. Really impressive. This can take in a ton of reference images and
merge them all together in just one video. So, here's an example of that. Or to make your character even more accurate, you can input multiple images
of a certain object or character at different angles, and then feed that into Bernini to generate a really consistent render of that character. Or
here's another example where we can input multiple photos of this scene at different angles and get
Bernini to generate an orbiting video of the scene. The awesome thing is this is out already. So, at the top of the page, if you click on this code
button and you scroll down a bit, here it contains all the instructions on how to download and run this locally on your computer. Plus, the ComfyUI
team is already trying to integrate this into ComfyUI. So, expect to have a decent workflow very soon.
Now, if you click on this Hugging Face page, note that this is based off of 1.2, which has two separate branches. There's a high branch and a low
branch, and you need a model for each one. Now, the total size of one of these branches is like 84 GB, which is really damn huge. I'm sure there's
going to be more quantized or compressed versions that would work on lower VRAM in the near future. So,
stay tuned for that. For now, if you're interested in reading further, I'll link to this main page in the description below. Also this week, NVIDIA
drops a really cool 3D reconstruction model called Deja View. This takes in multiple images of a scene, and it can reconstruct the entire scene in 3D.
So, here's an example of the NVIDIA headquarters. Here's another example where we can reconstruct
this canyon like this. So, here are some other examples for your reference where the input is multiple images and the output is this 3D Gaussian splat
with the camera positions along all the images. Here's another example for your reference. The clever part is how small and efficient this is.
Normally, these 3D reconstruction transformers get bigger by stacking a ton of layers together. Deja Vu
takes a different approach. It reuses the same transformer block over and over. Almost like the model is thinking through the scene in repeated
passes. Now, if you compare this to another top competitor called Depth Anything Three, which is way larger at like 10 times the size, you can see
that this new NVIDIA Deja Vu at only 117 million parameters is able to match its performance while using way
less compute. Here are some more quantitative metrics and as you can see, Deja Vu, which is in green, is the most efficient and the most performant.
The awesome thing is they've released this already. So, if you click on this code button and you scroll down a bit, here it contains the instructions
on how to download and run this locally on your computer. They've also released the code on how to
train and evaluate this. Note that at 100 million parameters, this is fairly tiny at only like 468 megabytes in size. If you're interested in reading
further, I'll link to this main page in the description below. Also this week, Google and Meta released a really cool project called Pager, which
stands for Panoramic Geometry Reconstruction. And this is a model for creating 360-degree panoramas. So,
instead of just taking a normal camera image, this actually takes in a single panoramic image and predicts the geometry of the whole surrounding
scene. The clever trick is that Pager doesn't build everything from scratch. It takes a 3D foundation model and adapts it to panoramas by treating the
six cube faces of a 360° image like a multi-view set and predicts the geometry of the whole surrounding
scene, including the depth and the normal estimation. Here are some quantitative benchmarks for your reference. If you compare this to other 360°
scene generators, you can see that Pedro on average outperforms the other competitors in terms of depth estimation as well as surface normal
estimation. The awesome thing is not only have they released the model to this, but they also released the data
sets. One is called Pano InfiniGen and this consists of like 70,000 panoramas from 20,000 indoor scenes and then also 7,000 outdoor urban panoramas.
They also released a Zurich Pano data set, which consists of 100 panoramic lidar scans across Zurich. So, at the top of the page, they've released the
code to this. If you click on this button and you scroll down a bit, here it contains all the
instructions on how to download and run these models on your computer. The Pedro model is fairly tiny at only 6 GB in size, so you should be able to
fit this in most consumer GPUs. And then at the top, they also released the two data sets over here. If you're interested in reading further, I'll
link to this main page in the description below. Also this week, Google releases a really cool
open-source real-time music generator. It's called Magenta Real-Time 2 and this is quite a big deal. It's designed to behave more like a real musical
instrument. So, instead of just typing a prompt like make me a jazz track and waiting for a file, you can actually control this live with MIDI, audio
and text and the model responds while you're actually playing.
[music]
[music]
[music]
[music]
[music]
[music]
[music]
So, this can take your musical input like notes from a MIDI keyboard, a style prompt, or an audio reference and outputs live generated music that
follows what you're doing. The most impressive part is latency. The first version of Magenta real time had around 3 seconds of delay, which is way too
slow if you want something that feels playable. But, this new version two brings it down to just 200
milliseconds. So, this is 2.4 billion parameters and you don't even need a GPU to run this. You can just run this on a MacBook. So, you know, instead
of just a standalone app that takes in a text prompt to generate a full song, Google is trying to make this usable inside actual music production
workflows, like a plugin in a DAW. And that's why this is really interesting. It's about making AI music
more controllable, responsive, and playable enough so that musicians can treat it as part of the creative process. So, at the top of this page, they
released the GitHub repo and you scroll down a bit. Here it contains all the instructions on how to download and run this locally on your computer. If
you're interested in reading further, I'll link to this main page in the description below. Also
this week, OpenAI is upgrading ChatGPT's memory system with something called dreaming. The simplest way to think about it is this. GPT is getting
better at keeping useful context fresh instead of relying on only saved notes. You see, older saved memories could work, but they might miss natural
context and become stale over time. So, dreaming is designed to synthesize memory from past conversations
in the background so ChatGPT can better understand your preferences, goals, and keep track of things that change over time. For example, let's say
you're planning a trip to Singapore. Without memory, it just gives you some generic stuff of the most popular tourist destinations, but with memory,
it'll produce more personalized responses. For example, if you're interested in things like wildlife
photography or quiet dinners or specific preferences for booking hotels. Now, here's the nice thing about this new dreaming feature. A few days later
after your trip, if you prompt it to find something for takeout that night, with the previous old stale memory, it might think you're still in
Singapore. So, it'll suggest some options in Singapore. But with this new updated dreaming memory, it
should be able to understand that your Singapore trip is now over and you're back home. So, it'll help you plan something that's actually close to
your current location. In terms of task accuracy, you can see that after adding this new dreaming feature, it improves the correctness by quite a lot.
Same with factual recall and staying correct over time as things change. The awesome thing is they are
rolling this dreaming feature out to free users over the coming weeks. And they're also going to increase memory capacity for paid users. If you're
interested in reading further, I'll link to this main page in the description below. Also this week, we have this project called Mama, which
completely removes the need for motion capture markers. This can basically take a multi-view video of people
moving together and turn that into accurate 3D human movements without needing any like marker mocap suits that you usually see in professional mocap
studios. So, this just takes a normal video taken from multiple camera angles. It analyzes all the movements from everyone in the scene, and it can
recreate a detailed 3D animation of every person, including their pose, shape, hands, and motion over
time. This really matters because multi-person motion capture is way harder than just a single person. When two people are hugging or dancing,
wrestling, making contact, the system has to figure out which part belongs to which person. There's also a ton of parts that are hidden and obscured.
So, Mama needs to kind of predict all of this and generate the movements of each person accurately. Now, to
train this model, the researchers built Mama Sin, which is a huge synthetic data set with 2.5 million crops covering extreme poses, hands, and complex
multi-person interactions. So, after training this Mama model on so much data, you can see its performance is very strong. On average, it outperforms
other competitors across all these benchmarks. The awesome thing is this is out already. So, at the
top of the page, if you click on this code button and you scroll down a bit, here it contains all the instructions on how to download and run this
locally on your computer. If you're interested further, I'll link to this main page in the description below. Also this week, we have not one but two
top image model releases. One of them is called Reeve 2, and this is incredibly performant. You can
generate really good-looking posters with accurate text, incredibly photorealistic images, as you can see here. And the strength of Reeve is that it
has this layout feature. So, here's an example of an image generation in action, and as it's generating, you can see that it first lays out the image
in a structured code-like form with bounding boxes across the image that determines where things
should be in the image. And then afterwards, it renders the final results with all the penguins in the right places. The useful part here is control.
Because the image has this intermediate structure, the system can understand where objects are, how text is placed, and how the composition works. So,
for example, it can generate this final image, but each section is actually separate. The bowl can
be a separate region from this cutting board, which is separate from this bowl of pasta, which is separate from this salad bowl. So, afterwards, you
can edit this further while keeping other elements intact. And then here are some additional examples. This is particularly good in generating posters
or infographics with a ton of text and different artistic designs. The impressive thing is, at least
according to the Arena text-to-image leaderboard, you can see that Reef 2 is actually number two, under GPT Image 2, but it beats Google's Nano Banana
2, which is already an impressive feat. Now, this is paid and closed source, but if you are interested in generating images where you can control the
composition of images like this, then Reef is currently one of the best options you can use. If
you're interested, I'll link to this main page in the description below. You've probably tried a bunch of different AI tools by now, but if you're not
sure which one to use for writing, coding, research, reasoning, image analysis, or everyday work, this is the ultimate bundle for you. I partnered
with HubSpot to create the AI Model Cheat Sheet Bundle. I put it in the description below for you to
access for free. Whether you're writing content, summarizing long documents, doing research, coding, or generating images, this bundle helps you
figure out which AI tool is most likely to give you the best results. It helps you stop experimenting randomly and start using the right AI model on
purpose. My favorite is the LLM selection cheat sheet I put together, which breaks down the current top AI
models in plain English. It shows you what each model is actually good at, where it falls short, and when you should use it. So, instead of wasting
time guessing whether you should use ChatGPT, Claude, Gemini, Perplexity, or another model, this cheat sheet helps you pick the right one for the
right task. You also get the task to model decision matrix, which maps common AI tasks best model for the
job. You can access all of this for free using the link in the description below. Thanks to HubSpot for partnering with me on this resource and
sponsoring this video. Now, in addition to Reef 2, we also have Ideogram 4 released this week, which is also one of the top image models out there.
The nice thing about this is they've actually released the models to this, so you can run it locally if you
prefer. Now, Ideogram has previously been really good in terms of generating topography, posters, and designs like this. And Ideogram 4 is even more
performant. It can generate a ton of readable text while keeping everything consistent and realistic. It can also generate a ton of different elements
and precise layouts, and it understands your prompt very well. One of my favorite features about
Ideogram is that you can generate transparent images. So, for example, you can swap out the background afterwards if you wish. This also generates
super realistic photos. It's also really good at understanding your prompt. Interestingly, how this works is instead of just a simple text prompt,
this actually works best if you input a prompt in JSON format. So, here's what the raw JSON format looks
like. For example, we can have a background like this, and then we can have different objects. For example, this object would be this canvas painting.
We'll have another object for this gray sofa, and another object for a small round side table. This is quite hard to read, so let me reformat it
better for you. It's basically like this, but in code format. Now, fortunately, you don't have to do
this manually. You can still just enter a simple text prompt and then plug it through an AI to help you rewrite it into JSON format. And you know, the
interesting thing is this kind of follows the same format as Reeve, where you can set bounding boxes across your canvas. And for each bounding box,
you can determine what exactly goes in that area. For example, here, we can choose to add a text flow
in this specific format. Or over here, we can choose to add this Golden Globe text. And then down here, we can choose to add a floating cat. So, you
know, this setup is pretty much the same as Reeve 2, where we can determine the composition of things in the image using these bounding boxes. And
like Reeve 2, the awesome thing about this is that each separate object in this image is a layer. So,
you can like edit each layer further while keeping the rest of the image intact. Exactly the same as Reeve. Here are some additional examples for your
reference. Like I said, Audiogram is particularly strong in terms of generating different topographies and posters, and even realistic images like
this. This is definitely among the best image models out there. The awesome thing is not only can you
try this out on their online platform, but they've also released the models for you to use offline. So, if you click on this GitHub repo, it contains
all the instructions for you to get started. Plus, ComfyUI has already released a workflow for this. So, if you want to use it using ComfyUI, I'll
also link to this page in the description below, where you can download the workflow and follow these
instructions to operate it. Now, while this is open weights and it's really good for certain things, there's one downside to this, which is it's
incredibly censored. So, if you try to generate any remotely spicy content, your generation is going to turn out like this. This is actually a
generated image from Ideogram with this text. This censorship seems to be baked into the model. You can't really
like bypass it effectively at the moment. So, that's the current downside of this model. But, I mean, if you want to create normal stuff like designs
or posters or just normal photos, then this is really good. So, in terms of this design Arena benchmark, Ideogram beats Nano Banana Pro and is just
slightly behind Nano Banana 2. Now, if you only look at the open weights models, then Ideogram is the
clear leader and by a huge margin. And while Nano Banana 2 or GPT image might beat Ideogram in terms of text rendering or prompt understanding,
spatial reasoning, the strength of Ideogram is this layout control. Where again, you can draw bounding boxes and control the exact composition of your
images. Now, if I look at this leaderboard by Arena, where people can blind test different AI models, you
can see that Reef 2 is all the way up here. As I mentioned previously, Ideogram is all the way down here at number nine. It doesn't seem to be as
performant as like Nano Banana, but it is the current top open-source model, at least according to this leaderboard. If you're interested in reading
further, I'll link to this main page in the description below. Also this week, Google releases a new
open-source model called Gemma 4 12B. This is what they call a unified encoder-free multimodal model, which is designed to be even more efficient.
This is designed to bring agentic AI abilities directly to your laptop. So, this sits in between the even smaller 4B variant, which is designed for
like running on just your phone, and the more advanced 26 billion parameter mixture of experts model. So,
this new 12B variant is like a medium-sized model that's actually way more efficient. First of all, this is multimodal, so this can take in text,
images, and now audio. Plus, it has reasoning and agentic behaviors. And the really interesting part is it's new architecture. So, it doesn't use any
multimodal encoders. Traditionally, if you plug in, you know, an image or audio, it needs to go through
an encoder step first to convert it into the right format to feed through the large language model. But here, they completely got rid of this step.
Both vision and audio inputs flow directly into the AI model. And if you look at some benchmark scores like general knowledge or agentic performance
as well as image analysis, you can see that Gemma 4 performs very closely with the 24B variant, which
is like twice the size. The awesome thing is this is small enough to run offline and locally with just 16 GB of VRAM or unified memory. And this is
released under the Apache 2 license, which has very minimal restrictions. You can even use this for commercial purposes. So, here's an example where
we can upload an audio file into Gemma 4 for it to output the transcript. And because it can understand
audio, we can also just speak to it with our voice instead of typing a prompt. >> Okay, this is transcribed pretty well. Let's use the voice edit
feature. This is powered by the new Gemma 4 12B model, and it works fully offline. Make it more exciting. Cool. Draft an email to our Google
Developers Community with this content. >> At the bottom here, it contains all the instructions on how you can
run it through all these LLM platforms like LM Studio or Ollama. And over here is where you can download the models. If you're interested in reading
further, I'll link to this main page in the description below. Also this week, Alibaba releases a new model, Qwen 3.7 Plus. This is a multimodal agent
model, which means it's built to combine vision and language into one system that can understand and
reason across different types of inputs. For example, you can give it text, images, video, or visual tasks, and it can reason through what it sees and
respond accordingly. This is also designed for agentic tasks, so you can easily plug this through Claude code or open claw or Quen code or other
agentic frameworks. This is like really good at autonomously doing long horizon tasks. So, for example,
here they got it to autonomously code for 11 hours to create a fully functional English vocabulary app. And here's the final result. Because it's able
to analyze the screen and the interface, it's able to actually test the app itself and then iterate further to improve or fix bugs. Or here's another
example where we can plug this into Claude code and also plug in some references of a stock
charting app and get it to code up this entire app from scratch. And here's what we get. It's able to also autonomously test this out itself. Now, if
you compare this with the original app, you can see it's very, very similar. Now, if you compare its performance across agentic coding and reasoning
and also world knowledge, you can see that on average it even beats similar competitors like DeepSeek
V4 Pro as well as GLM 5.1, Gemini 2.6. They also compared it with Opus, although it's like two generations behind. This is only 4.6 and not 4.8. Here
are some additional benchmarks across multimodal reasoning and visual coding. The important thing here is the direction that Quen is moving. The new
Quen models are becoming less like just a simple language model that answers simple questions and
more like general purpose assistants that can look at interfaces, understand visual context, and help with real-world workflows. Now, currently this
Quen 3.7 plus is only available through their Cloud Studio and via API. If you're interested in reading further, I'll link to this main page in the
description below. And Nvidia is not done cooking. So, this week they also released Cosmos 3. This is a
foundation model for physical AI, which means a model for robots, autonomous vehicles, and anything that needs to understand the real world and act in
it. The key here is that Cosmos combines a ton of different inputs together, including text, image, video, audio, and action. And it uses all of these
inputs to understand what's happening and predict what could happen next, and then generate videos
or actions of specific things. So, for example, here is a clip completely generated by Cosmos 3. You can see how this can be used to generate a ton of
synthetic training data to train autonomous driving algorithms. Or this could also be used to create a ton of synthetic data to train robots on how to
manipulate different types of objects. The awesome thing is this is fully open source, so they're
releasing the models, the training scripts, deployment tools, as well as the data sets for this. In terms of the model size, if you scroll down and
click on this Hugging Face link, note that there are two different variants of this. There's a super model, which is larger and better quality, and
then there's a nano version. The nano version is still pretty huge, so the total size of everything is
35 GB. The transformer itself is 30 GB in size, whereas the larger super model is around 130 GB. On this page, here's the code on how you can train
this yourself. Plus, they've also released the data sets that were used to train this, which is fantastic. So, this includes data sets with a ton of
robotics videos, physical interactions, and even like real-world simulations, autonomous driving, and
warehouse scenes. If you're interested in reading further, I'll link to this main page in the description below. Now, probably the most exciting
announcement from Nvidia for me as a consumer is the RTX Spark. This is a new superchip designed for laptops and small desktops. So, Nvidia isn't just
targeting large enterprise data centers, they're also trying to bring AI into smaller Windows PCs.
Instead of needing a giant workstation, this RTX Spark chip combines Nvidia AI and graphics into just one chip aimed at gamers, developers, and people
that want to run local AI agents. So, this includes a Blackwell GPU up to over 6,100 cores, up to a 20-core ultra-efficient CPU, up to a petaflop of
FP4 AI performance. And this is the most important part, up to 128 GB of unified memory. This
matters because local AI models are often limited by memory, not just raw speed. For example, if you want to run one of the best medium-sized models
locally, Qwen 3.6 35B, this has already gotten like over 5 million downloads. Notice that the total size of this is like 72 GB. Well, it's hard to fit
all of this in like a low-end consumer GPU. But with 128 GB of unified memory, this makes it a piece
of cake. So, now with such a laptop, you can easily run local AI models including image generators, video models, and of course, LLMs completely
offline. You don't need to purchase any subscription or connect to any cloud service. Now, these laptops are planned to go out in fall of this year,
so stay tuned for that. If you're interested in reading more, I'll link to this main page in the
description below. Also this week, Stability AI, which is the team behind Stable Diffusion, teases a new tool called Stable Layers. This can basically
turn an image into transparent layers like this. For example, if our input image is this, well, you can plug it through this AI to break it down into
these separate layers. And if you merge them all together, it becomes pretty much the same image.
Or here's another example. Now, the problem with training such a model is it usually needs paired data, which means you need the final image as well
as the different transparent layers for that image. And this data set is really hard to create. So, what they did is they completely skipped this
step. Instead, they just used another vision language model to score its generation and then through
reinforcement learning, eventually it learns how to separate the layers better and more accurately. And if you compare this new stable layers method
with another competitor called Kwen Image Layered, you can see that this new method, on average, performs better. It has better layer separation,
fewer errors, and cleaner alpha masks. It's also better at handling occluded areas. So, if you're looking
to convert an image into separate transparent layers, which you can edit further, this is one of the best options you can use. Now, currently they've
only released a technical paper on this. They do have a code button, but you can't click on it yet. Hopefully, they will open source this in the near
future. For now, if you're interested in reading further, I'll link to this main page in the
description below. In humanoid robot news, we have a new demo from the DR02 humanoid robot from the company Deep Robotics. In addition to Engine AI
and Unitree, this is another robotics company to put on your radar. So, DR2 is positioned to be an industrial-grade, all-weather humanoid robot built
for difficult environments. This has improved payload capacity, better obstacle crossing, and it's
reported to have like extreme dust and water protection. It can operate in extreme temperatures, so it's meant to support outdoor and harsh
conditions. As you can see here, this robot has the raw power and speed to do a sprint and flip, but it also has fine motor control and visual
accuracy to flip a delicate circuit breaker switch without crushing it. It's also able to successfully navigate
across different environments like steep hills or ledges, concrete stairs, so it's designed for unstructured and challenging environments. Now, this
demo from DR02 is pretty impressive, but you know what's even more exciting is this new demo from UBTech. If you can't get a partner in real life,
don't worry because your full-body humanoid companion is coming soon. So, UBTech just released this demo
teasing their emotional humanoid robot couple, which is planned to debut later this month. These are super realistic-looking robots. Now, I featured a
ton of these realistic-looking models before in the past weeks, but all of these are just head from this demo from UBTech it seems like they built a
full-body humanoid robot that looks super realistic. Also this week, one of my favorite AI labs
MiniMax releases their latest model MiniMax M3. This will be an open-source model with three incredible things at once. It has state-of-the-art
performance in terms of agentic coding plus a massive 1 million token context window, which means you can fit a ton of things into your prompt at
once, roughly 700,000 words or a medium-sized code base. Plus, it's also multimodal. So, this can support
image and video and even desktop actions as input. So, because of these abilities, it's able to reason through really long projects that require
multiple steps. It can understand visual information or even operate computer workflows. For example, if you get it to code an app, it can also then
pull up and look at the interface and verify that everything actually works. The interesting part is this
new MiniMax sparse attention, which basically lets the model look across huge amounts of context without the cost exploding. The benchmark scores are
also super strong. For these agentic coding benchmarks, especially for SweetBench Pro, it even beats GPT-5.5. The awesome thing is this will be
open-source in the coming 1 or 2 weeks. Now, because this will be open-source, what I'm curious about is
how it compares to the top open-source models out there including Kimmy, DeepSeek-V4, and JLM-5.1. If I look at this leaderboard by artificial
analysis, then it does look like MiniMax M3 is the leading open model, just one point above Kimi K2.6 and Mimo. And if you look at the cost, if you
choose to use it through their API, it's also very cheap and efficient. So, it only costs like 20 cents per
million tokens, which is similar to DeepSeek and Mimo, whereas Kimi K2 is like over three times more expensive. However, if I look at another
leaderboard called Arena, where people can blind test different models side by side, then we get a slightly different picture. So, in terms of overall
text responses, you can see that GLM 5.1 is the current leading open model, whereas MiniMax M3 is all the
way down here. Not too impressive. Now, if you look at this Web Dev Arena, then as you can see, GLM 5.1 is still the current leader, whereas MiniMax
M3 is a few points below it. But it still beats Kimi K2.6. Now, if you look at this other benchmark called the Valves Index, which looks at
performance across finance and coding tasks, then you can see MiniMax M3 does perform quite well above DeepSeek
V4 and Kimi K2.6. Now, if you look at its design capabilities, then MiniMax is an improvement compared to the previous version, but still slightly
behind GLM 5.1 and Kimi. So, overall, some mixed results depending on which benchmark or leaderboard you look at. But on average, it's still a very
capable state-of-the-art model. The awesome thing is they plan to release the model weights to this
sometime in the next one or two weeks. So, stay tuned for that. Right now, if you want to try this out, you can use it through their API or through a
ton of other third-party providers. If you're interested in reading further, I'll link to this main page in the description below. Also, this week,
what I think is an even more exciting update from Microsoft is this new chip, and actually the
headline isn't just the chip itself, but how much AI has helped build it. Basically, Microsoft used its discovery AI system to help manage workflows,
automate measurements, spot flaws that were easy to miss, and suggest new solutions during the chip development process. The result is a chip with a
new material stack that Microsoft says is a thousand times more reliable than the previous generation
of qubits. This matters because quantum computers are extremely sensitive, and reliability is one of the biggest things holding them back. Now, if
you're new to quantum computers, here's why they matter. These are meant to solve some real-world problems that are so complex that even today's
supercomputers struggle with. For example, simulating molecules and atoms and predicting how materials
behave to solving certain math problems, which can become insanely difficult as the problem gets bigger. But, quantum computers are designed to solve
these extremely hard problems much faster than even the best supercomputers today. So, these can be used for things like discovering new medicines or
designing better batteries, improving materials, or even breaking and improving encryption. And
Microsoft says it now expects to reach a scalable quantum computer by 2029. That's just 3 years away, which is super exciting. They basically cut
their previous timeline in half. So, actually the big story here is with the help of Project AI, they were able to accelerate the development of this
quantum chip. It's able to help scientists move faster through experiments, measurements, and design
decisions inside a real hardware pipeline. A really fascinating read. If you're interested in reading further, I'll link to this main page in the
description below. Also this week, ByteDance drops a new zero-shot text-to-speech generator called Wave TTS. This can instantly clone anyone's voice
with just a few seconds of their recording and make them speak out anything using a text transcript. So,
here are some examples. First of all, here's like 3 to 4 seconds of their original voice. >> I would have won the Junior Olympics if not for those
meddling kids. >> And then here's the transcript. Here's the generation from Wave TTS. >> Looking through the telescope, I saw a circle of deep blue
and the little round planet. >> Here are some other examples. I'll play you the original recording first
and then the generation. >> This man looked exactly the same except that now the roles were reversed. He also tried to remember some good stories to
relate as he shared the sheep. >> He had come to the town only to find a woman who could interpret his dream. Before the boy could reply, a butterfly
appeared and fluttered between him and the old man. >> The awesome thing is they released this
already. So on this page, if you scroll down a bit, it contains all the instructions on how to download and run this locally on your computer. They
also provide the script on how to train this and fine-tune it if you want. Note that this Wave TTS is based off of F5 TTS, which I've already covered
on my channel. This is one of the best TTS models out there. So if you want to learn more about that,
see this video for a full installation tutorial. Anyways, back to Wave TTS. Notice that the checkpoint is 10 GB in size, so this should be able to fit
in most consumer GPUs. If you're interested in reading further, I'll link to this main page in the description below. Next up, Alibaba drops a really
cool AI called Stream Care. This just takes in a prompt and a transcript and it's able to stream
that person in real time speaking out the transcript. So here are some examples. >> Guys, you literally have to try this makeup pen that feels like it
was designed by angels. For anyone who's ever struggled with shaky hands because the moment you glide this micro precision tip across your skin, you
realize it's not just an eyeliner or brow pencil, but a total game changer that instantly upgrades
your look from just woke up to red carpet ready.
So here's what the interface could look like. So, you can set up an avatar and then chat with it and it can respond almost in real time.
It's also able to follow precise motion control. For example, you can input a motion sequence like this plus a prompt and a transcript and here's the
result. >> Well, well, you know, these days I've just been kind of taking things one step at a time. Like, nothing too crazy. Just trying to get
through work, make time for family, maybe catch up with a few friends here and there. And honestly, I
feel like that's already enough sometimes because life can get pretty busy, right? So, I'm just trying to keep things >> The awesome thing is this can
also generate quite long videos, even exceeding 5 minutes. And the character and the overall scene remains consistent the whole time. Now, here it
says that Stream Character can run in real time on a single H100 GPU, which I'm sure most of you don't
have. For a consumer GPU, I wouldn't expect it to be able to run exactly in real time. There's probably some additional latency, although it still
should be faster than traditional video models you may be used to. Now, for now they've only released a technical paper on this. There's no indication
whether they'll release the code to this. If you're interested in reading further, I'll link to this
main page in the description below. Also this week, Nvidia releases Omni Dreams. This is a real-time generative world model for autonomous driving
simulation. Basically, it creates photorealistic multi-camera driving videos that react to the car's actions. The input can include a starting frame,
a text prompt, map information, and trajectory poses. And the output is a very realistic driving video.
In fact, it's super flexible. For example, you can generate videos with extreme weather like this. Or you can even generate some really specific and
strange situations like a mattress on top of a car falling down on the highway. You can even edit the video, for example, removing objects on the road
or adding a jaywalker or changing the color of a certain vehicle. Or here's an example where you can
input map information and a planned trajectory to get OmniDream to reconstruct a really realistic driving video that aligns with the inputs. So, this
is a very valuable tool for creating synthetic training videos for autonomous driving models. One of the biggest problems or bottlenecks with training
autonomous driving is we just simply don't have enough real-world data. But if we can get AI to
generate more videos that are physically accurate and realistic, then these autonomous driving models can learn way faster and more efficiently. The
awesome thing is they released OmniDreams already. So, if you click on this code button and you scroll down a bit, here it contains all the
instructions on how to download and run this locally on your computer. If you're interested in reading further,
I'll link to this main page in the description below. Also this week, Nvidia drops a beast. It's called Nemotron 3 Ultra and this is their best
open-source model. Here are the specs. This is quite a massive 550 billion parameter mixture of experts model. Think of this as like a team of AIs
working together to help you solve a task. And then when you use it, only 55 billion parameters are active.
So, this is quite efficient. Now, this is designed for agentic systems, so workflows that can plan, call tools, read files, delegate subtasks, verify
outputs, and keep working across many turns. Now, these workflows can get expensive really quickly because every step creates more tokens and more
context. But Nemotron tries to solve that with a ton of clever tweaks, like this hybrid Mamba
transformer architecture for long context, as well as this NVFP4 quantization for faster inference, as well as multi-token prediction to generate
things faster. Now, this has a 1 million token context window, which means you can fit a ton of information into it at once. And this can achieve five
times faster inference compared to the other leading open models. So, the X axis is the output speed.
Notice how fast this is. It's also super efficient. So, it takes around 30% lower in cost to complete the same task compared to other leading open
models. Now, if you click on their Hugging Face at over 500 billion parameters, this model is quite huge at 350 GB in size. You'll need to stack like
multiple GPUs to run this. If you're interested in reading further, I'll link to this main page in the
description below. Also this week, we have yet another text-to-speech generator called Higgs Audio version 3. And this is also super flexible and
capable. The most interesting feature here is control. You can add inline tags directly inside the prompt to dictate the emotion or speaking style,
speed, pitch, pauses, or even sound effects while the sentence is spoken. Here are some examples. >> Hey,
how can I help you? Same voice, same words, and a completely different presence. Welcome to Higgs Audio V3. I was built for chat native voice. Real-time, expressive, and controllable.
I can be happy. I also can be sad. I can slow things down. Or quickly speed things up. >> And as you can see here in terms of its performance against
other leading open source TTS models, you can see that Higgs audio overall performs the best. Now, if you scroll all the way down, they have released
the models on Hugging Face. So, if you click on this, here it contains more instructions on how to
use it, such as which tags it can support. They have a ton of emotion tags you can use, as well as different speech styles and sound effects and
prosody tags. The model is fairly small at only 9.3 GB in size, so you should be able to fit this on most mid-tier GPUs. If you're interested in
reading further, I'll link to this main page in the description below. Also this week, the Ernie team from
Baidu releases Neva. This is a new video generation framework with audio baked in. As you may know, one of the best open source models out there is
Alibaba's 1.2.2. However, that model by itself can only generate silent videos. What Neva tries to do here is add audio into the mix as well. So, here
are some examples. >> You can talk? >> Yes, with Neva, even winter magic comes alive. >> Small models
can still do big things. >> Something was wrong in this neighborhood. It started last week when the same cars kept returning after midnight. >> Can I
use any timbre I want? >> Yes, in all kinds of complex scenes. >> Here's how it works. Instead of just making a video and then trying to attach the
sound afterwards, it takes a prompt and generates the video with synchronized audio, including speech,
sound, timing, and motion. They describe it as a native audio-visual alignment, which means the model first builds a shared space where sound and
visuals can understand each other. And then it uses the prompt or context to guide the final result. Here it says this can do really fast generations
of 720p videos with dual channel audio and precise timbre control. And it also supports different aspect
ratios. The impressive thing here is that this is only 6.3 million parameters, but it seems to even beat the leading open model with audio, LT X2.3,
which is 19 billion parameters. The awesome thing is they released this already. So, if you click on this GitHub repo and you scroll down a bit here,
it contains all the instructions on how to download and run this locally on your computer. Nav a
model is 25 GB, also an FP8 version, which is only 7 GB. So, this should be able to fit in most consumer GPUs. If you're interested in reading
further, I'll link to this main page in the description below. Also this week, we have some exciting releases from Microsoft. So, it seems like
they're also starting to train their own AI models. So, one of them is called My, which stands for Microsoft AI
thinking one. This is a medium-sized thinking model. So, here are some specs. This has a trillion total parameters. This is a mixture of experts
models. So, at any given time when you use it, only 35 billion of these are active. So, it's pretty efficient. Now, given its size at only 35 billion
active parameters, it's actually pretty performant. So, if you compare this with Sonnet 4.6, it does
outperform it across several benchmarks. However, it's still not as performant as the other top closed and open-source models out there. So, I'm not
going to spend too much time on it. Here they say that this My thinking one is currently available in private preview in Microsoft Foundry, and it'll
also be available publicly soon. They also released My Image 2.5, which currently if you look at this
Arena leaderboard where people can blind test different AI models side by side, you can see that My Image actually beats Google's Nano Banana models.
And it ranks just behind GPT Image 2. So, much like Nano Banana and GPT Image 2, this My Image is also an image generator and editor. So, not only can
you just create images using a text prompt, but this can also edit images using natural language.
Here are some examples for your reference. It's quite performant. It can definitely generate a variety of different images. Now, currently this is
paid and closed source. It's only available in Microsoft Foundry, and it's not better than GPT image 2. So, I'm not going to spend too much time on
it. But, if you're interested in reading further, I'll link to this main release page in the description
below. And that sums up all the highlights in AI this week. Let me know in the comments what you think of all of this. Which piece of news was your
favorite, and which tool are you most looking forward to trying out? As always, I will be on the lookout for the top AI news and tools to share with
you. So, if you enjoyed this video, remember to like, share, subscribe, and stay tuned for more
content. Also, there's just so much happening in the world of AI every week. I can't possibly cover everything on my YouTube channel. So, to really
stay up-to-date with all that's going on in AI, be sure to subscribe to my free weekly newsletter. The link to that will be in the description below.
Thanks for watching, and I'll see you in the next one.