[Music] Apple just unveiled fast VLM, a vision language model that's 85 times faster, three times smaller, and powerful enough to run smoothly on a MacBook Pro. This could be the breakthrough that finally makes AI truly see and understand the world in real time. Let's start with why this matters in the first place. VLMs are the class of models that let AI systems handle both text and images together. So instead of just asking a model to write you an essay, you can show it a chart, a diagram, a scan document, even a screenshot, and it's supposed to understand and respond to you in a meaningful way. But here's the catch. Resolution is everything. If your input image is low resolution, the model misses details. But if you crank up the resolution, you run into a different problem. Suddenly, your vision encoder has to handle way more data. It outputs way more tokens, and that slows everything down. both on the vision side and on the language model side. And when you're sitting there waiting for that first token to appear on screen, that lag is what we call TTFT or time to first token. That number matters and Apple's been obsessed with it in this project. Over the years, researchers have tried all sorts of tricks to make vision language models faster and smarter. Some of the earliest systems like Frozen and Florence relied on a method called cross attention. Basically weaving together the image and text inside the layers of the language model so they could interact more closely. Later on the field shifted toward auto reggressive approaches, things like lava and plug owl, mini GPT4, Cambrian one, and plenty of others. These models don't bother weaving image and text together in the middle. Instead, they just feed the image information right alongside the text, letting the language model process both streams at once. The go-to engine for handling images in most of these systems has been clip style transformers. Variants like SIGLIP, AVA clip, intern VIT, and DFN clip are everywhere because they're reliable and well tested. But they come with a catch. They pump out a huge number of visual tokens which bogs down the rest of the system. To fix that, people have experimented with pruning, cutting down the token count on the fly. Approaches like lava primemerge or matrioska token sampling were designed specifically for this. At the same time, another group of researchers started moving away from flat transformer designs and toward what's called hierarchical backbones. With setups like convex or fastfitit, the image gets progressively downsampled in stages so you end up with fewer denser tokens and avoid being overwhelmed by raw data. And not too long ago, a model called conlava came out that ditched transformers altogether for the vision side, relying purely on convolutions. Apple looked at all of this, the pruning tricks, the hybrid models, the full convolutional encoders, and decided to go one step further. That's where Fast VM comes in built around a component they call Fast Vit HD. At its core, Fast Vit HD is a hybrid vision encoder, meaning it mixes convolutional layers with transformer layers. The convolutional part does the heavy lifting of compressing the image, pulling out local details quickly and efficiently. The transformer part then takes over handling the bigger picture, making sure the relationships between different parts of the image are preserved and refined before passing the information on to the language model. So, you're hearing about this AI opportunity that could build incredible wealth. But you're probably thinking, how do I actually seize it and turn it into real money? Here's what I can tell you. The method I discovered, the one that's generated over $500,000 in the past 12 months, is called Faceless Empire. It's the complete system for building automated income streams that work while you sleep. No need to show your face on video. No complicated tools, just AI doing the heavy lifting. But here's the thing, I'm only sharing this for a few days with 200 founding members. If you're someone who likes to seize opportunities instead of watching from the sidelines, click the link in the description to be the first to know once we reveal everything. Sign up for free right now to be notified when we reveal the whole system. The brilliance of Fast Vit HD is that it produces far fewer tokens than traditional approaches. Yet, it doesn't lose the fine details that matter for highresolution images, and it's built to keep latency, the time you wait before the model starts generating words, as low as possible. To understand why this matters, think about how most hybrid models like Fast Vit work. They use four stages to gradually shrink the image. Apple added a fifth stage with an extra downsampling layer. So instead of the self attention mechanism operating on data reduced by a factor of 16, it now works on data reduced by 32. That single adjustment means the encoder runs faster and ends up producing four times fewer tokens for the language model to process. The architecture is carefully balanced up front. There are three stages built around rep mixer blocks which are highly efficient for convolutional processing. After that come two stages powered by multi-headed self- attention which bring in the transformer's ability to reason over the entire image. Together, this setup delivers both efficiency and accuracy, fewer tokens, less lag, and no major sacrifice in detail. Numbers tell the story better than anything. In the Lava 1.5 setup, Fast VM gets a 3.2 2 times improvement in TTFT against Lava 1 Vision at its maximum resolution of 1152 by 1152. Fast VLM runs 85 times faster on TTFT with a vision encoder that's 3.4 times smaller. And these aren't just speed tricks. It's also matching or beating the benchmarks. On text VQA, it's 8.4% better than convoc. And across a broad set of benchmarks like Seedbench, MMU, and MMVET, it's either matching or surpassing state-of-the-art models, often while generating five times fewer visual tokens. When they compare it against Cambrian 1, which is a pretty strong multimodal model using multiple encoders, fast VLM is nearly eight times faster. The training setup itself is surprisingly efficient, too. Apple trained these models on a single node with eight Nvidia H180 GPUs. Stage one of training that only takes around 30 minutes with a Quen 27B decoder. Stage 1.5 where they do resolution scaling with 15 million samples is longer. About 77 hours if you're training at 1024 resolution. Stage two, the visual instruction tuning with 1.1 million samples takes about 8 hours. And if you scale further with 12 1.5 million highquality instruction tuning samples, you get even more performance gains. Apple also released multiple checkpoints, R4, R12, R41, and so on, so the community can dig in and test different training stages. Now, one of the clever things about this work is how they approached scaling instead of doing fancy token pruning or tiling strategies. They realized you could just scale the input resolution directly thanks to how fast Vid HD is designed. So rather than chopping up your image into tiles like any res or sphinx and then processing each tile independently, fast VLM can just take the full highresolution image, encode it efficiently and still keep latency down. They did test tiling though and found that at extreme resolutions like 1536 x 1536, tiling can still help, but in most cases direct scaling was better for both accuracy and latency. The design of Fast Vid HD is surprisingly lightweight. It has far fewer moving parts than the giant models out there. Yet, it runs circles around them in speed. On a regular MacBook Pro, it's nearly seven times faster than some of the most popular vision transformers while keeping accuracy just as high. Even against models built specifically for this kind of task like Vitamin, it's not only faster, but also smaller, which means it's more efficient all around. Apple also tested how well speed and accuracy balance out, and the results are clear. If you crank up resolution but pair it with a weak language model, the system wastes time. But with the right balance, say medium resolution and a solid LLM, you get both speed and precision. And compared to previous models, FastFit HD consistently gives better answers without the lag. What makes this even more exciting is that Apple ran all these tests on actual consumer hardware, not some massive server farm. They converted the model so it could run directly on the Mac neural engine and it performed brilliantly. That shows this isn't just a research toy. This is tech that could realistically run on your own devices in the future. And while other teams tried complicated tricks to cut down on the number of tokens these systems generate, Apple's design just does it naturally. At lower resolutions, it spits out as few as 16 tokens and still outperforms competitors that produce 10 times more. In other words, they solved the problem at the root instead of patching it later. The flexibility is another win. Whether you pair it with smaller language models like Quinn 2.5B or bigger ones like Quinn27B, it still shines. Even with the tiny LLMs, it beats older models by a huge margin, sometimes more than 80 times faster. With the bigger LLMs, it competes with the giants of the field while using just a fraction of the resources. And in benchmarks that focus on tough textheavy tasks. Things like OCR, document understanding, and chart analysis. Fast VLM pulls ahead. Competing models with billions of parameters often need thousands of tokens to get the job done. Fast VLM gets the same or better results with just over a hundred tokens. That's the kind of leap that makes this model not just a little better, but a whole new standard for how efficient AI can be. Apple's Fast Vit HD shows that hybrid vision encoders are the smarter path for multimodal AI. Convolutions bring speed and efficiency. Transformers add reasoning, and together they deliver better results without hacks like pruning or tiling. By benchmarking it on a MacBook Pro instead of giant servers, Apple made it clear this tech is built for real devices, hinting at future AI assistants that run locally and efficiently. So, there it is. Some will become millionaires because of AI. Some will stay exactly where they are, and some will unfortunately lose their jobs. Which one will you be? Faceless Empire gives you the exact method that's generated over $500,000 for us over the last 12 months. But only 200 people will get a chance to get our system when I reveal everything in a few days. Don't be the person who had the chance to seize this AI opportunity and didn't take it. Sign up for free priority access now. The link's below, but not for long. All right, that's it for now. If you enjoyed this breakdown, drop a comment, hit that like button, and make sure you're subscribed for more deep dives into what's happening in AI and robotics. Thanks for watching, and I'll catch you in the next one.